Computing device

By integrating audio and image data processing using neural networks to analyze speaker feature vectors, the method addresses the challenges of voice control in vehicles, enhancing accuracy and reliability in noisy environments.

JP2024161380AActive Publication Date: 2024-11-19SOUNDHOUND INC
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
JP2024114800
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2024-07-18
Publication Date
2024-11-19
Estimated Expiration
2039-08-23

AI Technical Summary

Technical Problem

Existing voice-controlled systems in vehicles face challenges due to limited processing resources, high noise levels, and complex acoustic environments, leading to difficulties in accurately transcribing and analyzing human speech, with current solutions being prone to errors and failures.

Method used

A method that combines audio and image data to process speech within vehicles, using a neural network architecture to analyze speaker feature vectors derived from facial areas, improving accuracy and robustness by incorporating visually derived information to compensate for noise and vehicle conditions.

Benefits of technology

Enhances the accuracy and reliability of speech processing in vehicles by leveraging speaker feature vectors from image data, reducing errors and improving responsiveness in noisy and resource-constrained environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024161380000001_ABST
    Figure 2024161380000001_ABST
Patent Text Reader

Abstract

To provide a device and method for analyzing human speech based on audio data and image data.SOLUTION: In one example, visual information is used to improve audio processing. This visual information can be image data obtained from inside a vehicle. In one example, a device includes a speech processing module configured to analyze human speech based on audio data and image data. The speech processing module includes an acoustic model configured to process the audio data and predict phoneme data used to analyze the speech. The acoustic model is configured to receive a speaker feature vector and audio data as input, and is trained to use the speaker feature vector and audio data to predict the phoneme data.SELECTED DRAWING: Figure 1B
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] FIELD OF THE PRESENT ART The present technology is in the field of audio processing. One example relates to processing audio captured from inside a vehicle. [Background technology]

[0002] background Recent advances in computing have made many long-desired voice-controlled applications more feasible. For example, improvements in statistical models, including practical frameworks for effective neural network architectures, have greatly increased the accuracy and reliability of previous voice processing systems. This, combined with the rise of wide-area computer networks, offers a range of modular services that can be accessed simply using application programming interfaces. Voice is rapidly becoming a viable option for providing user interfaces.

[0003] While voice control devices have become commonplace in the home, providing voice processing in a vehicle presents additional challenges. For example, vehicles often have limited processing resources for auxiliary functions (such as voice interfaces), experience significant noise (e.g., high levels of road noise and / or engine noise), and present constraints in the acoustic environment. Any user interface is further constrained by the safety implications of controlling the vehicle. These factors make voice control difficult to practically achieve in a vehicle.

[0004] Moreover, despite advances in speech processing, even users of advanced computing devices often report that current systems lack human-level responsiveness and intelligence. It is incredibly difficult to translate pressure fluctuations in the air into analyzed commands. Speech processing typically involves complex processing pipelines, and errors at any stage will undermine the success of the machine's interpretation. Many of these difficulties are not immediately obvious to humans, who can process speech without conscious thought using cortical or subcortical structures. However, engineers working in the field are rapidly realizing the gap between human capabilities and current state-of-the-art speech processing.

[0005] US8,442,820B2 describes a lip-reading and speech recognition combined multimodal interface system. The system can issue navigation operation commands only by voice and lip movement, which allows the driver to see the front during navigation operation and reduces vehicle accidents related to navigation operation while driving. The lip-reading and speech recognition combined multimodal interface system described in US8,442,820B2 includes an audio voice input unit, a speech recognition unit, a speech recognition command and estimated probability output unit, a lip video image input unit, a lip-reading unit, a lip recognition command output unit, and a speech recognition and lip recognition result combined unit that outputs the speech recognition command. Although US8,442,820B2 provides a solution for in-vehicle control, the proposed system is complex and the many interacting components increase the chances of errors and analysis failures.

[0006] A speech processing system that transcribes and analyzes human speech more accurately It is desirable to provide a speech processing method and system that can be implemented in a real-world device, such as an embedded computing system for a vehicle. Achieving a practical speech processing solution is difficult because of the many system integration and connectivity challenges that exist in vehicles. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] U.S. Patent No. 8,422,820 Summary of the Invention [Means for solving the problem]

[0008] Summary of the Invention Some examples described herein provide a method and system for processing speech. Some examples use both audio data and image data to process speech. Some examples are adapted to address the challenges of processing utterances captured in vehicles. Some examples obtain a speaker feature vector based on image data featuring at least a facial area of ​​a person, e.g., a person in a vehicle. Then, speech processing is performed using visually derived information that depends on the speaker of the utterance. This can improve accuracy and robustness.

[0009] In one aspect, an apparatus for a vehicle includes an audio interface configured to receive audio data from an audio capture device located in the vehicle, an image interface configured to receive image data from the image capture device characterizing a facial area of ​​a person in the vehicle to capture image data in the vehicle, and a voice processing module configured to analyze an utterance of the person based on the audio data and the image data. The voice processing module includes an acoustic model configured to process the audio data and predict phoneme data used to analyze the utterance, the acoustic model including a neural network architecture. The apparatus further includes a speaker pre-processing module implemented by the processor, the speaker pre-processing module configured to receive the image data and obtain a speaker feature vector based on the image data, the acoustic model configured to receive the speaker feature vector and the audio data as inputs and trained to use the speaker feature vector and the audio data to predict the phoneme data.

[0010] In the above aspect, a speaker feature vector is obtained using image data that features the face area of ​​the person speaking. This speaker feature vector is provided as an input to the neural network architecture of the acoustic model, and the acoustic model is configured to use this input as well as audio data that features the utterance. This provides the acoustic model with additional visually derived information that the neural network architecture can use to improve the analysis of the utterance, for example to compensate for undesirable acoustic and noise characteristics in the vehicle. For example, configuring the acoustic model based on a particular person and / or their mouth area determined from the image data can improve the determination of ambiguous phonemes that may otherwise be incorrectly transcribed based on the vehicle context, for example.

[0011] In one variation, the speaker pre-processing module is configured to perform facial recognition on the image data to identify a person in the vehicle and extract a speaker feature vector associated with the identified person. For example, the speaker pre-processing module may include a facial recognition module used to identify a user speaking in the vehicle. If the speaker feature vector is determined based on the audio data, the identification of the person may enable a predetermined (e.g., pre-computed) speaker feature vector to be extracted from the memory. This may reduce the limitation of The processing latency for some embedded vehicle control systems may be improved.

[0012] In one variation, the speaker pre-processing module includes a processor-implemented lip-reading module configured to generate one or more speaker feature vectors based on lip movements within a person's facial area, which may be used together with or independent of a facial recognition module, where the one or more speaker feature vectors provide a representation of the speaker's mouth or lip area that may be used by the neural network architecture of the acoustic model to improve processing.

[0013] In some cases, the speaker pre-processing module may include a neural network architecture configured to receive data derived from one or more of the audio data and image data and predict speaker feature vectors. For example, this approach may combine a vision-based neural lip-reading system with an acoustic "x-vector" system to improve processing of acoustics. If one or more neural network architectures are used, they may be trained using a training set that includes image data, audio data, and a ground truth set of linguistic features, such as a ground truth set of phoneme data and / or text transcriptions.

[0014] In some cases, the speaker pre-processing module is configured to calculate a speaker feature vector for a predefined number of utterances and to calculate a static speaker feature vector based on the plurality of speaker feature vectors for the predefined number of utterances. For example, the static speaker feature vector may include an average of a set of speaker feature vectors that are linked to a particular user using image data. The static speaker feature vector may be stored in a memory of the vehicle, which may also improve speech processing capabilities in resource-limited vehicle computing systems.

[0015] In one case, the device includes a memory configured to store one or more user profiles. In this case, the speaker pre-processing module is configured to perform facial recognition on the image data to identify a user profile in the memory associated with a person in the vehicle, calculate a speaker feature vector for the person, store the speaker feature vector in the memory, and associate the stored speaker feature vector with the identified user profile. Facial recognition can provide a quick and easy mechanism to extract useful information (e.g., speaker feature vector) for processing of specific person-dependent acoustics. In one case, the speaker pre-processing module can be configured to determine whether the number of stored speaker feature vectors associated with a given user profile is greater than a predefined threshold. If this is the case, the speaker pre-processing module may calculate a static speaker feature vector based on the number of stored speaker feature vectors, store the static speaker feature vector in memory, associate the stored static speaker feature vector with a given user profile, and indicate that the static speaker feature vector is to be used for future utterance analysis in place of calculating a speaker feature vector for the person.

[0016] In one variation, the apparatus includes an image capture device configured to capture electromagnetic radiation having infrared wavelengths, and the image capture device is configured to send image data to an image interface. This can provide an illumination-invariant image that enhances image data processing. The speaker pre-processing module can be configured to process the image data to extract one or more portions of the image data, and the extracted one or more portions are used to obtain a speaker feature vector. For example, the one or more portions can relate to a face area and / or a mouth area.

[0017] In one case, one or more of the audio interface, image interface, voice processing module, and speaker pre-processing module may be located in the vehicle and may include, for example, parts of a local embedded system. In this case, the processor may be located in the vehicle. In another case, the voice processing module may be remote from the vehicle. In this case, the device may include a transceiver that transmits data derived from the audio and image data to the voice processing module and receives control data from the analysis of the utterance. Different distributed configurations are possible. For example, in one case, the device may be realized locally in the vehicle, but yet another copy of at least one component of the device may be executed on a remote server device. In this case, certain functions may be executed remotely, for example, together with or instead of local processing. The remote server device may have improved processing resources that improve accuracy, but may increase processing latency.

[0018] In one case, the acoustic model includes a hybrid acoustic model including a neural network architecture and a Gaussian mixture model, the Gaussian mixture model configured to receive the vector of class probabilities output by the neural network architecture and output phoneme data for analyzing the utterance. The acoustic model additionally or alternatively includes, for example, a Hidden Markov Model (HMM) along with the neural network architecture. In one case, the acoustic model may include a connectionist temporal classification (CTC) model or another form of a neural network model having a recurrent neural network architecture.

[0019] In one variation, the speech processing module includes a language model communicatively coupled to the acoustic model to receive the phoneme data and generate a transcription representative of the utterance. In this variation, the language model may be configured to use, for example, a speaker feature vector in addition to the acoustic model to generate a transcription representative of the utterance. This may be used to improve language model accuracy when the language model includes a neural network architecture, such as a recurrent neural network or a transformer architecture.

[0020] In one variation, the acoustic model includes a database of acoustic model configurations, an acoustic model selector that selects an acoustic model configuration from the database based on a speaker feature vector, and an acoustic model instance that processes the audio data, the acoustic model instance being instantiated based on the acoustic model configuration selected by the acoustic model selector, and the acoustic model instance being configured to generate phoneme data used to analyze the utterance.

[0021] In an example, the speaker feature vector is one or more of an i-vector and an x-vector. The speaker feature vector may include a composite vector, for example, including two or more of a first speaker-dependent portion generated based on audio data, a second speaker lip movement-dependent portion generated based on image data, and a third speaker face-dependent portion generated based on image data.

[0022] According to another aspect, there is a method of processing an utterance, the method including receiving audio data from an audio capture device located within the vehicle, the audio data characterizing an utterance of a person within the vehicle, receiving image data from an image capture device located within the vehicle, the image data characterizing a facial area of ​​the person, obtaining a speaker feature vector based on the image data, and analyzing the utterance using a speech processing module implemented by a processor. Analyzing the utterance includes receiving the speaker feature vector as an input to an acoustic model of the speech processing module. The method includes providing a speech recognition system that recognizes speech based on a speaker feature vector and audio data, the acoustic model including a neural network architecture, and analyzing the utterance further includes predicting phoneme data using at least the neural network architecture based on the speaker feature vector and the audio data.

[0023] The method may provide a similar improvement to voice processing in a vehicle. In one case, obtaining a speaker feature vector includes performing facial recognition on the image data to identify a person in the vehicle, obtaining user profile data for the person based on the facial recognition, and obtaining a speaker feature vector according to the user profile data. The method may further include comparing a number of stored speaker feature vectors associated with the user profile data to a predefined threshold. In response to the number of stored speaker feature vectors being below the predefined threshold, the method may include calculating a speaker feature vector using one or more of the audio data and the image data. In response to the number of stored speaker feature vectors being greater than the predefined threshold, the method may include obtaining a static speaker feature vector associated with the user profile data, the static speaker feature vector being generated using the number of stored speaker feature vectors. In one case, the speaker feature vector includes processing the image data to generate one or more speaker feature vectors based on lip movement in a facial area of ​​the person. Analyzing the utterance may include providing phoneme data to a language model of the voice processing module, predicting a transcript of the utterance using the language model, and using the transcript to determine control commands for the vehicle.

[0024] According to another aspect, there is a non-transitory computer readable storage medium having instructions stored thereon that, when executed by at least one processor, cause the at least one processor to receive audio data from an audio capture device and receive a speaker feature vector derived based on image data from an image capture device, the image data characterizing a facial area of ​​a user, and further cause the at least one processor to analyze the utterance using a speech processing module, where analyzing the utterance includes providing the speaker feature vector and the audio data as input to an acoustic model of the speech processing module, the acoustic model including a neural network architecture, and further includes predicting phoneme data using at least the neural network architecture based on the speaker feature vector and the audio data, providing the phoneme data to a language model of the speech processing module, and generating a transcript of the utterance using the language model.

[0025] At least one processor may be configured to remotely, for example, from a motor vehicle. The computer-aided speech recognition system may include a computing device, such as a computing device, in which audio data and a speaker image vector are received from a motor vehicle. The instructions may enable the processor to perform automatic speech recognition with a lower error rate. In some cases, the speaker feature vector includes one or more of a speaker-dependent vector element generated based on the audio data, a speaker lip movement-dependent vector element generated based on the image data, and a speaker face-dependent vector element generated based on the image data. [Brief description of the drawings]

[0026] [Figure 1A] 1 is a schematic diagram showing the interior of a vehicle according to an example. [Figure 1B] 1 is a schematic diagram illustrating an apparatus for a vehicle according to an example. [Diagram 2]FIG. 1 is a schematic diagram illustrating an apparatus for a vehicle having a speaker pre-processing module according to an example. [Diagram 3] FIG. 2 is a schematic diagram illustrating components of a speaker pre-processing module according to an example. [Figure 4] FIG. 2 is a schematic diagram illustrating components of an audio processing module according to an example. [Diagram 5] FIG. 2 is a schematic diagram illustrating a neural speaker pre-processing module and a neural speech processing module according to an example. [Figure 6] FIG. 2 is a schematic diagram illustrating components that make up an acoustic model of a speech processing module according to an example. [Figure 7] FIG. 2 is a schematic diagram of an image pre-processor according to an example. [Figure 8] FIG. 2 is a schematic diagram illustrating image data from different image capture devices according to an example. [Figure 9] FIG. 2 is a schematic diagram illustrating components of a speaker pre-processing module configured to extract lip features according to an example. [Figure 10A] 1 is a schematic diagram illustrating a motor vehicle having an apparatus for voice processing according to an example. [Figure 10B] 1 is a schematic diagram illustrating a motor vehicle having an apparatus for voice processing according to an example. [Figure 11] FIG. 2 is a schematic diagram illustrating components of a user interface for a motor vehicle according to an example. [Figure 12] FIG. 1 is a schematic diagram illustrating an example computing device for a vehicle. [Figure 13] FIG. 1 is a flow diagram illustrating a method for processing an utterance according to an example. [Figure 14] FIG. 1 is a schematic diagram illustrating a non-transitory computer readable storage medium according to an example. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0027] Detailed Description Various examples of the present technology that demonstrate various interesting aspects are described below. In general, the examples may use the described aspects in any combination.

[0028] Certain examples described herein use visual information to improve voice processing. This visual information may be obtained from inside the vehicle. In examples, the visual information features a person in the vehicle, e.g., a driver or a passenger. Certain examples use the visual information to generate a speaker feature vector for use by an adapted voice processing module. The voice processing module may be configured to use the speaker feature vector to improve processing of associated audio data, e.g., audio data derived from an audio capture device in the vehicle. Examples may improve the responsiveness and accuracy of an in-vehicle voice interface. Certain examples may be used by a computing device to improve voice transcription. Thus, the described examples may be seen to extend a voice processing system with multimodal capabilities that improve the accuracy and reliability of voice processing.

[0029] Certain examples described herein provide different approaches to generate speaker feature vectors. Some approaches are complementary and can be used together to synergistically improve speech processing. In one example, image data obtained from inside the vehicle, e.g., from driver and / or passenger cameras, is processed to identify a person and determine a feature vector that numerically represents certain characteristics of the person. These features may include audio features, e.g., numerical representations of expected variations in audio data for an acoustic model. In another example, image data obtained from inside the vehicle, e.g., from driver and / or passenger cameras, is processed to determine a feature vector that numerically represents certain visual features of the person, e.g., features associated with vocalizations by the person. In one case, the visual features are associated with the person's mouth area, e.g., representing the position and / or movement of the lips. In both examples, the speaker feature vectors may have a similar format, so that the phoneme data may be ... phoneme data may be processed to determine a feature vector that numerically represents certain visual features of the person, e.g., features associated with vocalizations by the person. In one case, the phoneme data may be processed to determine a feature vector that numerically represents certain visual features of the person, e.g., features associated with vocalizations by the person. It can be easily integrated into the input pipeline of the acoustic models used to generate it. One example may provide an improvement that overcomes some of the challenges of automatic speech recognition in vehicles, such as the confined interior of the vehicle, the possibility of multiple people speaking within this confined interior, and high levels of engine and environmental noise.

[0030] Example Vehicle Context FIG. 1A illustrates an exemplary context for an audio processing device. In FIG. 1A, the context is an automotive vehicle. FIG. 1A is a schematic diagram of an interior 100 of an automotive vehicle. The interior 100 is shown on the front driver's side of the automotive vehicle. A person 102 is shown present within the interior 100. In FIG. 1A, the person is the driver of the automotive vehicle. The driver faces forward in the vehicle and observes the road through a windshield 104. The person controls the vehicle using a steering wheel 106 and observes vehicle status indications via a dashboard or instrument panel 108. In FIG. 1A, an image capture device 110 is located within the interior 100 of the automotive vehicle near the bottom of the dashboard 108. The image capture device 110 has a field of view 112 that captures a facial area 114 of the person 102. In this example, the image capture device 110 is positioned to capture an image through an aperture in the steering wheel 106. FIG. 1A also illustrates an audio capture device 116 located within the interior 100 of the automotive vehicle. The audio capture device 116 is arranged to capture sounds emitted by the person 102. For example, the audio capture device 116 may be arranged to capture sounds from the person 102, i.e., sounds emanating from the person's facial area 114. The audio capture device 116 is shown mounted on the windshield 104. For example, the audio capture device 116 may be mounted near or on the rearview mirror, or on a door frame against the side of the person 102. FIG. 1A further illustrates an audio processing device 120. The audio processing device 120 may be mounted on the motor vehicle. The audio processing device 120 may include or form part of a control system for the motor vehicle. In the example of FIG. 1A, the image capture device 110 and the audio capture device 116 are communicatively coupled to the audio processing device 120, for example, via one or more wired and / or wireless interfaces. The image capture device 110 may be located outside the motor vehicle to capture images within the motor vehicle through a window pane of the motor vehicle.

[0031] The context and configuration of FIG. 1A is provided as an example to aid in understanding the following description. It should be noted that the example need not be limited to autonomous vehicles, but may be implemented in other forms of vehicles as well. Other forms of vehicles include, but are not limited to, marine vehicles such as boats and ships, aerial vehicles such as helicopters, planes and gliders, rail vehicles such as trains and trams, spacecraft, construction vehicles and heavy equipment. Autonomous vehicles may include, for example, cars, trucks, sport utility vehicles, motorbikes, buses, and motorized carts. The use of the term "vehicle" herein further includes certain heavy equipment that may remain stationary and be motorized, such as cranes, lifting equipment, and boring equipment. Vehicles may be manually controlled and / or have autonomous capabilities. Although the example of FIG. 1A features a steering wheel 106 and a dashboard 108, other control configurations may be provided (e.g., an autonomous vehicle may not have a steering wheel 106 as shown). Although a driver's seat context is shown in FIG. 1A, a similar configuration may be provided for one or more passenger seats (e.g., both front and rear). FIG. 1A is provided for illustrative purposes only and omits certain features that may be present in a motor vehicle for clarity. In some cases, the approaches described herein may be used outside the context of a vehicle and may be implemented by a computing device, such as, for example, a desktop or laptop computer, a smartphone, or an embedded device.

[0032] FIG 1B is a schematic diagram of the audio processing device 120 shown in FIG 1A. In FIG 1B, the audio processing device 120 includes an audio processing module 130, an image interface 140, and an audio interface 150. The image interface 140 is configured to receive image data 145. The image data may include image data captured by the image capture device 110 in FIG 1A. The audio interface 150 is configured to receive audio data 155. The audio data 155 may include audio data captured by the audio capture device 116 in FIG 1A. The audio processing module 130 is communicatively coupled to both the image interface 140 and the audio interface 150. The audio processing module 130 is configured to process the image data 145 and the audio data 155 to generate a set of linguistic features 160 usable to analyze the speech of the person 102. Linguistic features may include phonemes, word parts (e.g., stems or root words), and words (including text features such as pauses that map to punctuation marks), as well as probabilities and other values ​​for these linguistic units. In one case, the linguistic features may be used to generate a text output representing the utterance. In this case, the text output may be used as is or may be mapped to a predefined set of commands and / or command data. In another case, the linguistic features may be directly mapped to a predefined set of commands and / or command data (e.g., without an explicit text output).

[0033] A person (such as person 102) may issue voice commands while operating an autonomous vehicle using the configuration of FIGS. 1A and 1B. For example, person 102 may speak internally, e.g., generate utterances, to control the autonomous vehicle or obtain information. An utterance in this context is associated with vocal sounds produced by a person that represent linguistic information such as speech. For example, an utterance may include sounds emanating from the larynx of person 102. An utterance may include a voice command, e.g., a spoken request from a user. For example, a voice command may include a request to perform an action (e.g., "play music," "turn on the air conditioning," "activate cruise control"), further information about the request (e.g., "album XY," "68 degrees Fahrenheit," "60 mph for 30 minutes"), transcribed speech (e.g., "add ... to my todo list" or "send the next message to user A"), and / or a request for information (e.g., "what's the traffic on C?", "what's the weather today?", or "where is the nearest gas station?").

[0034] The audio data 155 may take a variety of forms depending on the implementation. In general, the audio data 155 may originate from time-series measurements from one or more audio capture devices (e.g., one or more microphones), such as the audio capture device 116 in FIG. 1A. In some cases, the audio data 155 may be captured from one audio capture device. In other cases, the audio data 155 may be captured from multiple audio capture devices, e.g., multiple microphones at different locations within the interior 100. In the latter case, the audio data may include one or more channels of time-correlated audio data from each audio capture device. The audio data at the time of capture may include, for example, one or more channels of Pulse Code Modulation (PCM) data at a predefined sampling rate (e.g., 16 kHz), with each sample represented by a predefined number of bits (e.g., 8 bits, 16 bits, or 24 bits per sample, where each sample includes an integer or floating-point value).

[0035] In some cases, audio data 155 may be processed (e.g., pre-processed for voice processing) after capture but before receipt at audio interface 150. Processing may include one or more of filtering in one or more of the time and frequency domains, applying noise reduction and / or normalization. In one case, the audio data may be converted to measurements over time in the frequency domain, for example by performing a Fast Fourier Transform to produce one or more frames of spectrogram data. In some cases, a Mel filter bank or Mel frequency filter may be used to determine values ​​for one or more frequency domain features. A filter bank, such as Mel-Frequency Cepstral Coefficients, may be applied. In these cases, the audio data 155 may include the output of one or more filter banks. In other cases, the audio data 155 may include time-domain samples and pre-processing may be performed within the audio processing module 130. Different combinations of approaches are possible. Thus, the audio data as received at the audio interface 150 may include any measurements made along the audio processing pipeline.

[0036] In a similar manner to audio data, image data described herein may take a variety of forms depending on the implementation. In one case, image capture device 110 may include a video capture device, and the image data includes one or more frames of video data. In another case, image capture device 110 may include a still image capture device, and the image data includes one or more frames of still images. Thus, image data may originate from both video and still image sources. References to image data herein may, for example, refer to image data originating from a two-dimensional array having a height and width (e.g., equivalent to the rows and columns of the array). In one case, the image data may have multiple color channels, for example, three color channels for each of the colors red, green, and blue (RGB), each color channel having an associated two-dimensional array of color values ​​(e.g., 8 bits, 16 bits, or 24 bits per array element). The color channels may be referred to as different image "planes." In some cases, only a single channel may be used, for example, representing a "gray" or brightness channel. Different color spaces may be used depending on the application. For example, an image capture device may natively generate a frame of YUV image data characterized by a brightness channel Y (e.g., luminance) and two opponent color channels U and V (e.g., two chrominance components roughly aligned blue-green and red-green). Similar to audio data 155, image data 145 may be processed after capture. For example, one or more image filtering operations may be applied and / or image data 145 may be resized and / or cropped.

[0037] With reference to the examples of FIGS. 1A and 1B, one or more of the image interface 140 and audio interface 150 may be local to hardware within the automotive vehicle. For example, the image interface 140 and audio interface 150 may each include a wired coupling of a respective image and audio capture device to at least one processor configured to implement the voice processing module 130. In one case, the image and audio interfaces 140, 150 may include a serial interface through which the image and audio data 145, 155 may be received. In a distributed vehicle control system, the image and audio capture devices 140, 150 may be communicatively coupled to a central system bus, and the image and audio data 145, 155 may be stored in one or more storage devices (e.g., random access memory or solid state storage). In this latter case, the image and audio interfaces 140, 150 may include a communicative coupling of at least one processor configured to implement the voice processing module to the one or more storage devices. For example, the at least one processor may receive the image and audio data 145, 155. 55 may be configured to read data from a given memory location. In some cases, image and audio interfaces 140, 150 may include wireless interfaces and voice processing module 130 may be remote from the motor vehicle. Different approaches and combinations are possible.

[0038] While FIG. 1A shows an example where the person 102 is a driver of a motor vehicle, in other applications, one or more image and audio capture devices may be arranged to capture image data featuring a person not controlling the motor vehicle, such as a passenger. For example, a motor vehicle may have multiple image capture devices arranged to capture image data relating to people present in one or more passenger seats of the vehicle (e.g., different locations in the vehicle, such as the front and rear). Audio capture devices may be arranged to capture vocalizations from different people as well, e.g., microphones may be located in each door or door frame of the vehicle. In one case, multiple audio capture devices may be provided in the vehicle, and audio data may be captured from one or more of these for providing the data to the audio interface 150. In one case, pre-processing of the audio data may include selecting audio data from a channel believed to be closest to the person making the vocalization and / or combining audio data from multiple channels in the motor vehicle. As described below, certain examples described herein facilitate voice processing in vehicles with multiple passengers.

[0039] Exemplary Speaker Pre-Processing Module 2 illustrates an exemplary voice processing device 200. For example, voice processing device 200 may be used to implement voice processing device 120 shown in FIGURES 1A and 1B. Voice processing device 200 may form part of an in-vehicle automatic speech recognition system. In other cases, voice processing device 200 may be adapted for use outside a vehicle, such as in a home or office.

[0040] The speech processing device 200 includes a speaker pre-processing module 220 and a speech processing module 230. The speech processing module 230 may be similar to the speech processing module 130 of FIG. 1B. In this example, the image interface 140 and the audio interface 150 are omitted for clarity, but they may form part of the image input for the speaker pre-processing module 220 and the speech input for the speech processing module 230, respectively. The speaker pre-processing module 220 is configured to receive the image data 245 and output the speaker feature vector 225. The speech processing module 230 is configured to receive the audio data 255 and the speaker feature vector 225 and use them to generate linguistic features 260.

[0041] The voice processing module 230 is realized by a processor. The processor may be a processor of a local embedded computing system in the vehicle and / or a processor of a remote server computing device (a so-called "cloud" processing device). In one case, the processor may include a piece of dedicated voice processing hardware, for example one or more Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs), and so-called "system on chip" (SoC) components. In another case, the processor may be configured to process computer program code, for example firmware, stored in an accessible storage device and loaded into memory for execution by the processor. The voice processing module 230 is configured to analyze the speech of a person, for example person 102, based on the audio data 225 and the image data 245. In the present case, the image data 245 is preprocessed by the speaker preprocessing module 220 to generate a speaker feature vector 225. Similar to the voice processing module 230, the speaker preprocessing module 220 may: It can be any combination of hardware and software. In one case, the speaker pre-processing module 220 and the voice processing module 230 can be implemented on a common embedded circuit board for the vehicle.

[0042] In one case, the speech processing module 230 includes an acoustic model configured to process the audio data 255 and predict phoneme data used to analyze the utterance. In this case, the linguistic features 260 may include phoneme data. The phoneme data may relate to one or more phoneme symbols, for example, from a predefined alphabet or dictionary. In one case, the phoneme data may include a predicted sequence of phonemes. In another case, the phoneme data may include probabilities for one or more of a set of phoneme components, such as phoneme symbols and / or subsymbols from a predefined alphabet or dictionary, and a set of state transitions (e.g., for a hidden Markov model). The acoustic model may be configured to receive audio data in the form of an audio feature vector. The audio feature vector may include numerical values ​​representing one or more of the Mel-frequency cepstral coefficients (MFCCs) and the filter bank output. In one case, the audio feature vector may relate to a current window in time (often referred to as a "frame") and include differences in feature changes between the current window and one or more other windows in time (e.g., previous windows). The current window may have a width in the range of w milliseconds, e.g., in one case, w may be about 25 milliseconds. Other features may include, for example, a signal energy metric and the output of logarithmic scaling. The audio data 255 after pre-processing may include a frame (e.g., vector) of multiple elements (e.g., 10 to over 1000 elements), each element including a numerical representation associated with a particular audio feature. In one example, about 25-50 mel filterbank features and a similarly sized set of intra features (e.g., , there may be a similarly sized set of delta features (representing first-order derivatives) and a similarly sized set of double-delta features (representing second-order derivatives).

[0043] The speaker preprocessing module 220 may be configured to obtain the speaker feature vector 225 in many different ways. In one case, the speaker preprocessing module 220 may obtain at least a portion of the speaker feature vector 225 from a memory, for example, via a lookup operation. In one case, a portion of the speaker feature vector 225, including the i and / or x vectors described below, may be extracted from the memory. In this case, the image data 245 may be used to determine a particular speaker feature vector 225 to extract from the memory. For example, the image data 245 may be sorted by the speaker preprocessing module 220 to select one particular user from a set of enrolled users. In this case, the speaker feature vector 225 may include a numerical representation of features correlated to the particular user selected. In another case, the speaker preprocessing module 220 may calculate a speaker feature vector 225. For example, the speaker preprocessing module 220 may calculate a compressed or dense numerical representation of salient information in the image data 245. This may include a vector having many elements that are smaller in size than the image data 245. In this case, the speaker pre-processing module 220 may realize an information bottleneck to compute the speaker feature vector 225. In one case, the computation is determined based on a set of parameters, such as a set of weights, biases, and / or probability coefficients. Values ​​for these parameters may be determined via a training phase using a set of training data. In one case, the speaker feature vector 225 may be buffered or stored as a static value after a set of computations. In this case, the speaker feature vector 225 may be extracted from memory based on the image data 245 during a subsequent utterance. Further examples illustrating how the speaker feature vector may be computed are described below. If the speaker feature vector 225 includes a lip movement component, this component may be provided on a real-time or near real-time basis, eliminating the need for data storage. cannot be extracted from the page.

[0044] In one case, the speaker feature vector 225 may include a fixed-length one-dimensional array (e.g., a vector) of numerical values, e.g., one value for each element of the array. In other cases, the speaker feature vector 225 may include a multi-dimensional array, e.g., two or more dimensions represent multiple one-dimensional arrays. The numerical values ​​may include integer values ​​(e.g., within a range set by a particular bit length (e.g., 8 bits gives a range of 0 to 255)) or floating point values ​​(e.g., specified as 32-bit or 64-bit floating point values). Floating point values ​​may be used if normalization is applied to the visual feature tensors, e.g., where values ​​are mapped to a range of 0 to 1 or -1 to 1. As an example, the speaker feature vector 225 may include a 256 element array, where each element is an 8 or 16-bit value, although the form may vary based on the implementation. In general, a speaker feature vector 225 has less information content than a corresponding frame of image data, e.g., using the example above, a speaker feature vector 225 of length 256 with 8-bit values ​​is smaller than a 640x480 video frame with three channels of 8-bit values, i.e., 2048 bits vs. 7372800 bits. Information content may be measured in bits or in the form of an entropy measure.

[0045] In one case, the speech processing module 230 includes an acoustic model, which includes a neural network architecture. For example, the acoustic model may include one or more of a Deep Neural Network (DNN) architecture with multiple hidden layers, a hybrid model including a neural network architecture and one or more of a Gaussian Mixture Model (GMM) and a Hidden Markov Model (HMM), and a Connectionist Temporal Classification (CTC) model, which includes, for example, one or more recurrent neural networks that operate on a sequence of inputs and generate a sequence of linguistic features as output. The acoustic model may include one or more of a Deep Neural Network (DNN) architecture with multiple hidden layers, a hybrid model including a neural network architecture and one or more of a Gaussian Mixture Model (GMM) and a Hidden Markov Model (HMM), and a Connectionist Temporal Classification (CTC) model that includes, for example, one or more recurrent neural networks that operate on a sequence of inputs and generate a sequence of linguistic features as output. It may output predictions at the frame level (e.g., for phoneme symbols or subsymbols) and may use previous (and in some cases future) predictions to determine possible or most likely sequences of phoneme data for an utterance. Approaches such as beam search and the Viterbi algorithm may be used to generate predictions at the output of the acoustic model. The acoustic model may be trained on a time-step by time-step basis, thereby further determining the sequence of phoneme data that is output from the acoustic model.

[0046] If the speech processing module 230 includes an acoustic model and the acoustic model includes a neural network architecture (e.g., a "neural" acoustic model), the speaker feature vector 225 may be provided as an input to the neural network architecture along with the audio data 255. The speaker feature vector 225 and the audio data 255 may be combined in many ways. In a simple case, the speaker feature vector 225 and the audio data 255 may be concatenated into a longer combined vector. In another case, different input pre-processing may be performed on each of the speaker feature vector 225 and the audio data 255. For example, one or more attention layers, feedforward layers, or the like may be used to pre-process the speaker feature vector 225 and the audio data 255. A feed-forward layer and / or an embedding layer may be applied, The results of these layers may then be combined. Different sets of layers may be applied to different inputs. In other cases, the speech processing module 230 may include another form of statistical model, e.g., a probabilistic acoustic model, and the speaker feature vector 225 includes one or more numerical parameters (e.g., probability coefficients) to configure the speech processing module 230 for a particular speaker.

[0047] The exemplary audio processing device 200 provides improvements for audio processing in a vehicle. In the vehicle, there may be high levels of environmental noise, such as road and engine noise. There may further be acoustic distortion caused by the enclosed interior space of the motor vehicle. These factors may make it difficult to process the audio data in the comparative example. For example, the voice processing module 230 may fail to generate linguistic features 260 and / or generate poorly matching sequences of linguistic features 260. However, the configuration of FIG. 2 allows the voice processing module 230 to be configured or adapted based on speaker features determined based on the image data 245. This provides additional information to the voice processing module 230, such that it may select linguistic features that are consistent with a particular speaker, for example by exploiting correlations between appearance and acoustic features. These correlations may be long-term temporal correlations, such as general facial appearance, and / or short-term temporal correlations, such as specific lip and mouth positions. This leads to higher accuracy despite challenging noise and acoustic contexts. This may help reduce speech analysis errors, for example, improving the end-to-end transcription path and / or improving the audio interface for making voice commands. In some cases, this example may be able to utilize an existing driver-facing camera that is typically configured to monitor the driver to check for drowsiness and / or distraction. In some cases, there may be a speaker-dependent feature vector component that is extracted based on a recognized speaker and / or a speaker-dependent feature vector component that includes mouth movement features. The latter component may be determined based on features that are not configured for an individual user, e.g., common features for all users may be applied, but mouth movements are associated with a speaker. In some other cases, the extraction of mouth movement features may be configured based on a particular identified user.

[0048] Facial Recognition Example FIG. 3 illustrates an exemplary speech processing device 300. The speech processing device 300 illustrates additional components that may be used to realize the speaker pre-processing module 220 in FIG. 2. Certain components illustrated in FIG. 3 are similar to their counterparts illustrated in FIG. 2 and have similar reference numbers. Features described above with reference to FIG. 2 may also apply to the example 300 in FIG. 3. Like the exemplary speech processing device 200 in FIG. 2, the exemplary speech processing device 300 in FIG. 3 includes a speaker pre-processing module 320 and a speech processing module 330. The speech processing module 330 receives audio data 355 and a speaker feature vector 325 and calculates a set of linguistic features 360. The speech processing module 330 may be configured in a manner similar to the example described above with reference to FIG. 2.

[0049] In FIG. 3, a number of subcomponents of the speaker preprocessing module 320 are shown. These include a face recognition module 370, a vector generator 372, and a data store 374. Although these are shown as subcomponents of the speaker preprocessing module 320 in FIG. 3, in other examples they may be implemented as separate components. In the example of FIG. 3, the speaker preprocessing module 320 receives image data 345 that features a facial area of ​​a person. The person may include a driver or a passenger in a vehicle as described above. The face recognition module 370 performs face recognition on the image data to identify a person, for example, a driver or a passenger in a vehicle. The face recognition module 370 may include any combination of hardware and software to perform face recognition. In one case, the face recognition module 370 may be implemented using off-the-shelf hardware components such as B5T-007001 supplied by Omron Electronics Corporation. In this example, the face recognition module 370 detects a user based on the image data 345 and outputs a user identifier 376. The user identifier 376 is passed to a vector generator 372. The vector generator 372 uses the user identifier 376 to obtain a speaker feature vector 325 associated with the identified person. In some cases, the vector generator 372 may extract the speaker feature vector 325 from a data store 374. The speaker feature vector 325 is then stored in the data store 374 as shown in FIG. The audio signal is passed to the audio processing module 330 for use as described with reference to FIG.

[0050] In the example of FIG. 3, the vector generator 372 may obtain the speaker feature vector 325 in different manners depending on a set of operating parameters. In one case, the operating parameters include a parameter indicating whether a certain number of speaker feature vectors 325 have been calculated for a particular identified user (e.g., as identified by the user identifier 376). In one case, a threshold value is defined that is associated with a certain number of previously calculated speaker feature vectors. If the threshold value is 1, the speaker feature vector 325 may be calculated for a first utterance and then stored in the data store 374, and for subsequent utterances, the speaker feature vector 325 may be extracted from the data store 374. If the threshold value is 1 or more, such as n, n speaker feature vectors 325 may be generated, and the (n+1)th speaker feature vector 325 may be obtained as a composite function of the previous n speaker feature vectors 325 extracted from the data store 374. The composite function may include an average or an interpolation. In one case, once the (n+1)th speaker feature vector 325 is computed, it is used as the static speaker feature vector for a configurable number of future utterances.

[0051] In the above example, run-time computational requirements for the in-vehicle system may be reduced by the use of a data store 374 that stores the speaker feature vectors 325. For example, the data store 374 may include a local data storage device in the vehicle, such that the speaker feature vectors 325 may be extracted from the data store 374 for a particular user, rather than being calculated by the vector generator 372.

[0052] In one case, at least one computational function used by vector generator 372 may involve a cloud processing resource (e.g., a remote server computing device), in which case, in situations where there is limited connectivity between the vehicle and the cloud processing resource, speaker feature vector 325 may be extracted from local storage as a static vector rather than relying on any functionality provided by the cloud processing resource.

[0053] In one case, the speaker pre-processing module 320 may be configured to generate a user profile for each newly recognized person in the vehicle. For example, prior to or upon detection of the speech, such as captured by an audio capture device, the face recognition module 370 may attempt to match the image data 345 against previously observed faces. If no match is found, the face recognition module 370 may generate (or direct the generation of) a new user identifier 376. In one case, a component of the speaker pre-processing module 320, such as the face recognition module 370 or the vector generator 372, may be configured to generate a new user profile if no match is found, and the new user profile may be indexed using the new user identifier. The speaker feature vector 325 may then be associated with the new user profile, and the new user profile may be stored in the data store 374, ready to be extracted if a future match is made by the face recognition module 370. Thus, an in-vehicle image capture device may be used for face recognition to select a user-specific voice recognition profile. The user profile may be calibrated through an enrollment process, for example, when a driver first uses the vehicle, or may be learned based on data collected during use.

[0054] In one case, the speaker processing module 320 may be configured to perform a reset of the data store 374. At the time of manufacture, there may be no user profile information present in the device 374. During use, a new user profile is created as described above, A user may command a reset of the stored user identifier. In some cases, the reset may only occur during professional service, such as when the vehicle is serviced at a service shop or sold by an authorized dealer. In some cases, the reset may be provided at any time, via a user-provided password.

[0055] In an example where the vehicle includes multiple image capture devices and multiple audio capture devices, the speaker pre-processing module 320 may provide further functionality to determine the appropriate face area from one or more captured images. In one case, audio data from multiple audio capture devices may be processed to determine the closest audio capture device associated with the utterance. In this case, the closest image capture device associated with the determined closest audio capture device may be selected, and image data 345 from this device may be sent to the face recognition module 370. In another case, the face recognition module 370 may be configured to receive multiple images from multiple image capture devices, each image including an associated flag to indicate whether it should be used to identify the currently speaking user. Thus, the voice processing device 300 of FIG. 3 may be used to identify a speaker from multiple people in the vehicle and configure the voice processing module 330 for the particular characteristics of that speaker. This may further improve voice processing in the vehicle when multiple people are speaking within the constrained interior of the vehicle.

[0056] iVector In certain examples described herein, a speaker feature vector, such as speaker feature vector 225 or 325, may include data generated based on audio data, such as audio data 255 or 355 in FIG. 2 and FIG. 3. This is indicated by a dashed line in FIG. 3. In one case, at least a portion of the speaker feature vector may include a vector generated based on factor analysis. In this case, the utterance may be represented as a vector M that is a linear function of one or more factors. The factors may be combined in a linear and / or nonlinear model. One of these factors may include a speaker and session independent supervector m, which may be based on a Universal Background Model (UBM). Another of these factors may include a speaker dependent vector w. This latter factor may further be channel or session dependent, or yet another factor may be provided that is channel and / or session dependent. In one case, the factor analysis is performed using a Gaussian Mixture Model (GMM) mixture. In a simple case, a speaker utterance may be represented by a supervector M determined as M=m+Tw, where T is a matrix that defines at least the speaker subspace. The speaker-dependent vector w may have multiple elements with floating-point values. The speaker feature vector in this case may be based on the speaker-dependent vector w. One way to calculate w, sometimes called the "i-vector", is described in the paper "Front-End Factor Analysis For Speaker Verification" by Najim Dehak, Patrick Kenny, Reda Dehak, Pierre Dumouchel and Pierre Ouellet.This paper was published in IEEE Transactions On Audio, Speech And Language Processing 19, no.4, pages 788-798, 2010, and is incorporated herein by reference. In one example, at least a portion of the speaker feature vector includes at least a portion of an i-vector. An i-vector can be viewed as a speaker-dependent vector determined for an utterance from audio data.

[0057] In the example of FIG. 3, the vector generator 372 may calculate the i-vector for one or more utterances. If there are no speaker feature vectors stored in the data store 374, The i-vectors may be calculated by the vector generator 372 based on one or more frames of audio data for the utterance 355. In this example, the vector generator 372 may repeat the i-vector calculations for each utterance (e.g., for each voice query) until a threshold number of calculations are performed for a particular user, as identified using a user identifier 376 determined from the facial recognition module 370, for example. In this case, after a particular user is identified based on the image data 345, an i-vector is stored in the data store 374 for the user for each utterance. The i-vectors are also used to output the speaker feature vector 325. Once a threshold number of calculations are performed, such as, for example, 100 or so i-vectors are calculated, the vector generator 372 may be configured to calculate a profile for the particular user using the i-vectors stored in the data store 374. The profile may use the user identifier 376 as an index and may include a static (e.g., unchanging) i-vector calculated as a composite function of the stored i-vectors. The vector generator 372 may be configured to calculate the profile upon receipt of the (n+1)th query or as part of a background or periodic function. In one case, the static i-vector may be calculated as an average of the stored i-vectors. Once a profile has been generated by the vector generator 372 and stored in the data store 374, for example using a user identifier to associate the profile with a particular user, it may be extracted from the data store 374 instead of calculating the i-vector for the user and used for future utterance analysis. This may reduce the computational overhead of generating speaker feature vectors and may reduce i-vector variance.

[0058] x vector In an example, a speaker feature vector, such as speaker feature vector 225 or 325, may be calculated using a neural network architecture. For example, in one case, the vector generator 372 of the speaker pre-processing module 320 of FIG. 3 may include a neural network architecture. In this case, the vector generator 372 may calculate at least a portion of the speaker feature vector by reducing the dimensionality of the audio data 355. For example, the vector generator 372 may include one or more deep neural network layers configured to receive one or more frames of the audio data 355 and output a fixed length vector output (e.g., one vector per language). One or more pooling, nonlinear function, and SoftMax layers may further be included. In one case, the speaker feature vectors may be provided in the paper "Spoken Language Recognition using X-vectors" by David Snyder, Daniel Garcia-Romero, Alan McCree, Gregory Sell, Daniel Povey, and Sanjeev Khudanpur. ) which was published in Odyssey (pp. 105-111) in 2018 and is incorporated herein by reference.

[0059] The x-vector may be used in a similar manner to the i-vector described above, and the above approach is applied to the speaker feature vector generated using the x-vector and the i-vector. In one case, both the i-vector and the x-vector may be determined, and the speaker feature vector may include a super-vector that includes elements from both the i-vector and the x-vector. Both the i-vector and the x-vector may be combined by concatenation or weighted sum, e.g., typically floating point values ​​and / or values ​​normalized within a given range. In this case, the data store 374 may include stored values ​​for one or more of the i-vectors and x-vectors, whereby once a threshold is reached, a static value is calculated and stored with a particular user identifier for future extraction. In one case, interpolation may be used to determine a speaker feature vector from one or more i-vectors and x-vectors. In one case, interpolation may be used to extract different speaker feature vectors from the same vector source. This can be done by averaging the following speaker feature vectors:

[0060] If the speech processing module includes a neural acoustic model, a fixed length format for the speaker feature vector may be defined. The neural acoustic model may then be trained using the defined speaker feature vector, for example as determined by speaker pre-processing module 220 or 320 in FIG. 2 and FIG. 3. If the speaker feature vector includes elements derived from one or more of the i-vector and x-vector calculations, the neural acoustic model may "learn" to configure acoustic processing based on speaker-specific information embodied or embedded in the speaker feature vector. This may increase acoustic processing accuracy, especially in vehicles such as motor vehicles. In this case, image data provides a mechanism to quickly associate a particular user with a calculated or stored vector element.

[0061] Exemplary Audio Processing Module Figure 4 illustrates an exemplary audio processing module 400. The audio processing module 400 may be used to implement the audio processing modules 130, 230 or 330 in Figures 1, 2 and 3. In other examples, other audio processing module configurations may be used.

[0062] Per the previous example, the speech processing module 400 receives audio data 455 and a speaker feature vector 425. The audio data 455 and the speaker feature vector 425 may be configured according to any of the examples described herein. In the example of FIG. 4, the speech processing module 400 includes an acoustic model 432, a language model 434, and an utterance parser 436. As described above, the acoustic model 432 generates phoneme data 438. The phoneme data may include one or more predicted sequences of phoneme symbols or subsymbols, or other forms of proto-language units. In some cases, In, multiple predicted sequences may be generated along with probability data indicating the likelihood of a particular symbol or sub-symbol at each time step.

[0063] The phoneme data 438 is communicated to a language model 434, e.g., the acoustic model 432 is communicatively coupled to the language model 434. The language model 434 is configured to receive the phoneme data 438 and generate a transcription 440. The transcription 440 may include text data, e.g., character strings, word parts (e.g., stems and endings), or words. The characters, word parts, and words may be selected from a predefined dictionary, e.g., a predefined set of possible outputs at each time step. In some cases, the phoneme data 438 may be processed before passing to the language model 434 or may be preprocessed by the language model 434. For example, beamforming may be applied to the probability distributions (e.g., for phonemes) output from the acoustic model 432.

[0064] The language model 434 is communicatively coupled to the utterance parser 436. The utterance parser 436 receives the transcription 440 and uses it to analyze the utterance. In some cases, the utterance parser 436 generates utterance data 442 as a result of analyzing the utterance. The utterance parser 436 may be configured to determine a command and / or command data associated with the utterance based on the transcription. In one case, the language model 434 may generate multiple possible text sequences, for example with probability information for units in the text, and the utterance parser 436 may be configured to determine a completed text output, for example in the form of an ASCII or Unicode character encoding or a voice command or command data. Once it is determined that the transcription 440 includes a voice command, the utterance parser 436 may be configured to execute the command according to the command data or to command the execution of the command. This may result in response data being output as the utterance data 442. The utterance data 442 may include responses, such as command instructions, that are relayed to the person making the utterance, and providing an output on the dashboard 108 and / or via the vehicle's voice system. In some cases, the language model 434 may include a statistical language model, and the utterance parser 436 may include a separate "meta" language model configured to rescore alternative hypotheses as output by the statistical language model. This may be via an ensemble model that uses voting to determine a final output, such as a final transcription or command identification.

[0065] 4 illustrates with solid lines an example of an acoustic model 432 receiving speaker feature vector 425 and audio data 455 as inputs and using the inputs to generate phoneme data 438. For example, the acoustic model 432 may include a neural network architecture (including hybrid models having other non-neural components), where the speaker feature vector 425 and audio data 455 may be provided as inputs to the neural network architecture, and the phoneme data 438 is generated based on the output of the neural network architecture.

[0066] The dashed lines in FIG. 4 indicate additional connections that may be configured in an implementation. In a first case, the speaker feature vector 425 may be accessed by one or more of the language model 434 and the utterance parser 436. For example, if the language model 434 and the utterance parser 436 further include respective neural network architectures, these architectures may be configured to receive the speaker feature vector 425 as an additional input, for example, in addition to the phoneme data 438 and the transcription 440, respectively. If the utterance data 442 includes a command identifier and one or more command parameters, the complete speech processing module 400 may be trained in an end-to-end manner given a training set with ground truth outputs and training samples for the audio data 455 and the speaker feature vector 425.

[0067] In a second implementation, the speech processing module 400 of FIG. 4 may include one or more recurrent connections. In one case, the acoustic model may include a recurrent model, such as an LSTM. In another case, there may be feedback between modules. In FIG. 4, a dashed line indicates a first recurrent connection between the utterance parser 436 and the language model 434, and a dashed line indicates a second recurrent connection between the language model 434 and the acoustic model 432. In this second case, the current state of the utterance parser 436 may be used to configure future predictions of the language model 434, and the current state of the language model 434 may be used to configure future predictions of the acoustic model 432. The recurrent connection may be omitted in some cases to simplify the processing pipeline and enable easier learning. In one case, the recurrent connection may be used to calculate an attention or weight vector to be applied at the next time step.

[0068] Neural Speaker Preprocessing Module 5 illustrates an example speech processing apparatus 500 that employs a neural speaker pre-processing module 520 and a neural speech processing module 530. In FIG. 5, the speaker pre-processing module 520, which may implement modules 220 or 320 in FIG. 2 and FIG. 3, includes a neural network architecture 522. In FIG. 5, the neural network architecture 522 is configured to receive image data 545. In other cases, the neural network architecture 522 may further receive audio data, such as audio data 355, as illustrated by the dashed pathway in FIG. 3. In these other cases, the vector generator 372 of FIG.

[0069] In FIG. 5, the neural network architecture 522 includes at least a convolutional neural network. In some architectures, there may be one or more feedforward neural network layers between the last convolutional neural network layer and the output layer of the neural network architecture 522. The neural network architecture 522 may include adapted forms of AlexNet, VGGNet, GoogLeNet, or ResNet architectures. The neural network architecture 522 may be replaced in a modular manner as more accurate architectures become available.

[0070] The neural network architecture 522 outputs at least one speaker feature vector 525, which may be derived and / or used as described in any of the other examples. FIG. 5 illustrates, for example, a case where the image data 545 includes multiple frames from a video camera, where the frames feature a human face area. In this case, multiple speaker feature vectors 525 may be calculated using the neural network architecture, for example, one for each input frame of image data. In other cases, there may be a many-to-one relationship between frames of input data and speaker feature vectors. Note that using a recurrent neural network system, it is not necessary to synchronize the samples of the input image data 545 and the output speaker feature vector 525 in time. For example, the recurrent neural network architecture may operate as an encoder (or integrator) over time. In one case, the neural network architecture 522 may be configured to generate an x-vector as described above. In one case, the x-vector generator may be configured to receive image data 545, process the image data using a convolutional neural network architecture, and then combine the output of the convolutional neural network architecture with the audio-based x-vector. In one case, known x-vector configurations may be extended to receive image data and audio data and generate a single speaker feature vector that embodies information from both modal pathways.

[0071] In FIG. 5, the neural speech processing module 530 is a speech processing module such as one of the modules 230, 330, 400 that includes a neural network architecture. For example, the neural speech processing module 530 may include a hybrid DNN-HMM / GMM system and / or a fully neural CTC system. In FIG. 5, the neural speech processing module 530 receives frames of audio data 555 as input. Each frame may correspond to a time window, such as a window of w milliseconds elapsed for time series data from an audio capture device. The frames of audio data 555 may be asynchronous with the frames of image data 545, e.g., the frames of audio data 555 are likely to have a higher frame rate. Also, retention mechanisms and / or recurrent neural network architectures may be applied within the neural speech processing module 530 to provide temporal encoding and / or integration of samples. In another example, the neural speech processing module 530 is configured to process the frames of audio data 555 and the speaker feature vector 525 to generate a set of linguistic features. As discussed herein, references to neural network architectures include one or more neural network layers (in one case, a "deep" architecture having one or more hidden layers as well as multiple layers), where each layer may be separated from the next by a nonlinearity such as a tanh unit or a rectified linear unit (RELU). Other functions may be embodied within the layers, including pooling operations.

[0072] The neural audio processing module 530 may include one or more components as shown in FIG. 4. For example, the neural audio processing module 530 may include at least one The neural network architectures of the neural speaker pre-processing module 520 and the neural speech processing module 530 may be trained jointly. In this case, the training set may include frames of image data 545, frames of audio data 555, and ground truth linguistic features (e.g., ground truth phoneme sequences, text transcriptions or voice command classifications, and command parameter values). Both the neural speaker pre-processing module 520 and the neural speech processing module 530 may be trained in an end-to-end manner using this training set. In this case, errors between the predicted linguistic features and the ground truth linguistic features may be propagated back through the neural speaker pre-processing module 520 after the neural speech processing module 530. Parameters for both neural network architectures may then be determined using a gradient descent approach, which results in: The neural network architecture 522 of the neural speaker pre-processing module 520 may "learn" parameter values ​​(such as values ​​for weights and / or biases for one or more neural network layers) that generate one or more speaker feature vectors 525 that improve at least acoustic processing in the in-vehicle environment, and the neural speaker pre-processing module 520 learns to extract features from human facial areas that help improve the accuracy of the output linguistic features.

[0073] Training of neural network architectures as described herein is typically not performed on an in-vehicle device (although it can be performed if desired). In one case, training may be performed on a computing device with access to significant processing resources, such as a server computing device having multiple processing units (whether CPU, GPU, Field Programmable Gate Array (FPGA) or other dedicated processor architecture) and a large memory portion that holds the training data set. In some cases, training may be performed using a coupled accelerator device, e.g., a combinatorial FPGA or GPU-based device. In some cases, the learned parameters may be downloaded to the in-vehicle device from a remote server device, e.g., as part of an over-the-air update. The signal may be communicated to the embedded system.

[0074] Acoustic model selection example FIG. 6 illustrates an example speech processing module 600 that uses speaker feature vectors 625 to configure an acoustic model. The speech processing module 600 may be used to at least partially realize one of the speech processing modules described in other examples. In FIG. 6, the speech processing module 600 includes a database of acoustic model configurations 632, an acoustic model selector 634, and an acoustic model instance 636. The database of acoustic model configurations 632 stores a number of parameters for configuring an acoustic model. In this example, the acoustic model instance 636 may include a general acoustic model that is instantiated (e.g., configured or calibrated) using a particular set of parameter values ​​from the database of acoustic model configurations 632. For example, the database of acoustic model configurations 636 may store multiple acoustic model configurations. Each acoustic model configuration may be associated with a different user and includes one or more default acoustic model configurations that are used when no user is detected or when a user is detected but not specifically recognized.

[0075] In some cases, the speaker feature vector 625 can be used to represent a particular regional accent on behalf of (or with) a particular user. This can be useful in countries such as India where many different regional accents may exist. In this case, the speaker feature vector 625 can be used to dynamically load an acoustic model based on the accent recognition performed using the speaker feature vector 625. For example, this can be done by This may be possible in the case where the speaker feature vector 625 includes an x-vector as described above. This may be useful when having multiple accent models (e.g., multiple acoustic model configurations for each accent) stored in the vehicle's memory. This may then allow multiple separately trained accent models to be used.

[0076] In one case, the speaker feature vector 625 may include a classification of the person in the vehicle. For example, the speaker feature vector 625 may be derived from the user identifier 376 output by the face recognition module 370 in FIG. 3. In another case, the speaker feature vector 625 may include a set of classifications and / or probabilities output by a neural speaker pre-processing module, such as module 520 in FIG. 5. In the latter case, the neural speaker pre-processing module may include a softmax layer that outputs "probabilities" for a set of potential users (including a classification for "unrecognized"). In this case, one or more frames of input image data 545 may result in a single speaker feature vector 525.

[0077] In FIG. 6, the acoustic model selector 634 receives a speaker feature vector 625, for example from a speaker pre-processing module, and selects an acoustic model configuration from a database of acoustic model configurations 632. This may be operated in a manner similar to the example of FIG. 3 described above. If the speaker feature vector 625 includes a set of user classifications, the acoustic model selector 634 may select an acoustic model configuration based on these classifications, for example by sampling a probability vector and / or selecting the largest probability value as the determined person. Parameter values ​​for the selected configuration may be extracted from the database of acoustic model configurations 632 and used to instantiate an acoustic model instance 636. Thus, different acoustic model instances may be used for different identified users in the vehicle.

[0078] 6, acoustic model instance 636, as configured by acoustic model selector 634 using configurations extracted from database of acoustic model configurations 632, also receives audio data 655. Acoustic model instance 636 is configured to generate phoneme data 660 for use in analyzing utterances associated with (e.g., featured within) audio data 655. Phoneme data 660 may include, for example, a sequence of phoneme symbols from a predefined alphabet or dictionary. Thus, in the example of FIG. 6, acoustic model selector 634 selects an acoustic model configuration from database 632 based on the speaker feature vector, and the acoustic model configuration is used to instantiate acoustic model instance 636 to process audio data 655.

[0079] The acoustic model instance 636 may include both neural and non-neural architectures. In one case, the acoustic model instance 636 may include a non-neural model. For example, the acoustic model instance 636 may include a statistical model. The statistical model may use symbol frequencies and / or probabilities. In one case, the statistical model may include a Bayesian model, such as a Bayesian network or classifier. In these cases, the acoustic model configuration may include a specific set of symbol frequencies and / or prior probabilities measured in different environments. Thus, the acoustic model selector 634 allows a specific source (e.g., person or user) of an utterance to be determined based on both visual (and in some cases audio) information, which may provide an improvement over using the audio data 655 alone to generate the phoneme sequence 660.

[0080] In another case, the acoustic model instance 636 may include a neural model. In this case, the acoustic model selector 634 and the acoustic model instance 636 may include a neural network architecture. In this case, the database of acoustic model configurations 632 may be The acoustic model selector 634 may be omitted and the acoustic model selector 634 may provide vector inputs to the acoustic model instance 636 to construct the instance. In this case, the training data may be constructed from the image data used to generate the ground truth set of speaker feature vectors 625, audio data 655, and phoneme output 660. Such a system may be trained jointly.

[0081] Exemplary Image Preprocessing 7 and 8 illustrate example image preprocessing operations that may be applied to image data obtained from inside a vehicle, such as an automotive vehicle. FIG. 7 illustrates an example image preprocessing pipeline 700 that includes an image preprocessor 710. The image preprocessor 710 may include any combination of hardware and software to implement the functionality as described herein. In one case, the image preprocessor 710 may include hardware components that form part of an image capture circuit that is coupled to one or more image capture devices, and in another case, the image preprocessor 710 may be implemented by computer program code (such as firmware) executed by a processor of an in-vehicle control system. In one case, the image preprocessor 710 may be implemented as part of a speaker preprocessing module described in the examples herein, and in other cases, the image preprocessor 710 may be communicatively coupled to a speaker preprocessing module.

[0082] In FIG. 7, the image pre-processor 710 receives image data 745, such as an image from the image capture device 110 in FIG. 1A. The image pre-processor 710 processes the image data to extract one or more portions of the image data. FIG. 7 shows an output 750 of the image pre-processor 710. The output 750 may include one or more image annotations, such as metadata associated with one or more pixels of the image data 745, and / or features defined using pixel coordinates within the image data 745. In the example of FIG. 7, the image pre-processor 710 performs face detection on the image data 745 to determine a first image area 752. The first image area 752 may be cropped and extracted as image portion 762. The first image area 752 may be defined using a bounding box (e.g., at least top left and bottom right (x,y) pixel coordinates for a rectangular area). Face detection may be a precursor step for face recognition. For example, face detection may determine face areas in the image data 745, and face recognition may classify the face areas as belonging to a given person (e.g., in a set of people). In the example of FIG. 7, the image pre-processor 710 further identifies a mouth area in the image data 745 to determine a second image area 754. The second image area 754 may be cropped and extracted as the image portion 764. The second image area 754 may also be defined using a bounding box. In one case, the first and second image areas 752, 754 may be determined in relation to a set of detected facial features 756. These facial features 756 may include one or more of the eye, nose, and mouth areas.Detection of the facial features 756 and / or one or more of the first and second areas 752, 754 may use a neural network approach or the Viola-Jones face detection algorithm as described in "Robust Real-Time Face Detection" by Paul Viola and Michael J Jones, International Journal of Computer Vision 57, pp. 137-154, The Netherlands, 2004, which is incorporated herein by reference. In one example, one or more of the first and second image areas 752, 754 are used by a speaker pre-processing module described herein to obtain a speaker feature vector. For example, the first image area 752 may provide input image data for the face recognition module 370 in FIG. 3 (i.e., may be used to supply image data 345). An example of using the second image area 754 is described below with reference to FIG. 9.

[0083] FIG. 8 illustrates the effect of using an image capture device configured to capture electromagnetic radiation having infrared wavelengths. In some cases, image data 810 obtained by an image capture device such as image capture device 110 in FIG. 1A may be affected by low light conditions. In FIG. 8, image data 810 includes areas of shadow 820 that partially obscure face areas (e.g., including first and second image areas 752, 754 in FIG. 7). In these cases, an image capture device configured to capture electromagnetic radiation having infrared wavelengths may be used. This may include providing adaptations (e.g., removable filters in hardware and / or software) to image capture device 110 in FIG. 1A, and and / or a Near-Infra-Red (NIR) camera. The output from such an image capture device is shown diagrammatically as image data 830. In the image data 830, the face area 840 is reliably captured. In this case, the image data 830 provides an illumination-invariant representation that is not affected by changes in illumination, such as those that may occur while driving at night. In these cases, the image data 830 is provided to the image pre-processor 710 and / or the speaker pre-processing module, as described herein.

[0084] lip reading example In some examples, the speaker feature vectors described herein may include at least a set of elements that represent the mouth or lip features of a person. In these cases, the speaker feature vector may be speaker dependent as it varies based on the content of image data that features the mouth or lip area of ​​the person. In the example of FIG. 5, the neural speaker pre-processing module 520 may encode the lip or mouth features that are used to generate the speaker feature vector 525. These may be used to improve the performance of the speech processing module 530.

[0085] 9 illustrates another exemplary speech processing apparatus 900 that uses lip features to form at least a portion of a speaker feature vector. As with the previous examples, the speech processing apparatus 900 includes a speaker pre-processing module 920 and a speech processing module 930. The speech processing module 930 receives audio data 955 (in this case frames of audio data) and outputs linguistic features 960. The speech processing module 930 may be configured as in other examples described herein.

[0086] In FIG. 9, the speaker pre-processing module 920 is configured to receive two different sources of image data. In this example, the speaker pre-processing module 920 receives a first set of image data 962 that features a person's face area. This may include the first image area 762 as extracted by the image pre-processor 710 in FIG. 7. The speaker pre-processing module 920 further receives a second set of image data 964 that features a person's lip or mouth area. This may include the second image area 764 as extracted by the image pre-processor 710 in FIG. 7. The second set of image data 964 may be relatively small, such as a small cropped portion of a larger image obtained using the image capture device 110 of FIG. 1A. In other examples, the first and second sets of image data 962, 964 may not be cropped and may include copies of sets of images from the image capture device. Different configurations are possible. That is, cropping the image data may provide improvements in processing speed and learning, but the neural network architecture may be trained to work for a wide variety of image sizes.

[0087] The speaker pre-processing module 920 includes two components in FIG. 9: a feature extraction component 922 and a lip feature extractor 924. The lip feature extractor 924 forms part of the lip reading module. The feature extraction component 922 is the same as the speaker pre-processing module in FIG. The feature extraction component 922 may be configured in a manner similar to that of the module 320. In this example, the feature extraction component 922 receives a first set of image data 962 and outputs a vector portion 926 consisting of one or more of an i-vector and an x-vector (e.g., as described above). In one case, the feature extraction component 922 receives a single image per utterance, and the lip feature extractor 924 and the audio processing module 930 receive multiple frames over the time of the utterance. In one case, if the facial recognition performed by the feature extraction component 922 has a confidence value below a threshold, the first set of image data 962 may be updated (e.g., by using another / current frame of the video) and the facial recognition is reapplied until the confidence value reaches the threshold (or exceeds a predefined number of attempts). As described with reference to FIG. 3, the vector portion 926 may be calculated based on the audio data 955 for the first number of utterances and is extracted as a static value from memory once the first number of utterances have been exceeded.

[0088] The lip feature extractor 924 receives a second set of image data 964. The second set of image data 964 may include cropped frames of image data focusing on the mouth or lip area. The lip feature extractor 924 may receive the second set of image data 964 at the frame rate of the image capture device and / or at a subsampled frame rate (e.g., every 2 frames). The lip feature extractor 924 outputs a set of vector portions 928. These vector portions 928 may include the output of an encoder that includes a neural network architecture. The lip feature extractor 924 may include a convolutional neural network architecture to provide a fixed length vector output (e.g., 256 or 512 elements with integer or floating point values). The lip feature extractor 924 may output a vector portion for each input frame of image data 964, and / or may extract features over time steps using a recurrent neural network architecture (e.g., using a Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU), or a “Transformer” architecture). In the latter case, the output of the lip feature extractor 924 may include one or more of a hidden state of a recurrent neural network and an output of a recurrent neural network. One exemplary implementation for the lip feature extractor 924 is described in "Lip reading sentences in the wild" by Chung, Joon Son et al., at the 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), which is incorporated by reference herein.

[0089] In FIG. 9 , the speech processing module 930 receives as input the vector portion 926 from the feature extraction component 922 and the vector portion 928 from the lip feature extractor 924. In one case, the speaker pre-processing module 920 may combine the vector portions 926, 928 into a single speaker feature vector. In another case, the speech processing module 930 may receive the vector portions 926, 928 separately but treat them as different portions of a speaker feature vector. The vector portions 926, 928 may be combined into a single speaker feature vector by one or more of the speaker pre-processing module 920 and the speech processing module 930, for example, using concatenation or more complex attention-based mechanisms. If the sample rates of one or more of the vector portions 926, 928 and the frames of audio data 955 are different, then the vector portions 926, 928 may be combined into a single speaker feature vector by one or more of the speaker pre-processing module 920 and the speech processing module 930, for example, using a receive-and-hold architecture (e.g., a new (slower, more varying values ​​are held constant at a given value until a sample value is received), recurrent temporal encoding (e.g. using LSTM or GRU as above), or attention weighting per time step. A common sample rate can be achieved by a vector-changing attention-based system.

[0090] The audio processing module 930 may be configured to use the vector portions 926, 928 as described in other examples described herein, e.g., they may be input into a neural acoustic model as speaker feature vectors along with the audio data 955. In examples in which the audio processing module 930 includes a neural acoustic model, a training set may be generated based on the input video from an image capture device, the input audio from an audio capture device, and the ground truth linguistic features (e.g., the image pre-processor 710 in FIG. 7 may be used to obtain first and second sets of image data 962, 964 from the raw input video).

[0091] In some examples, vector portion 926 may further include an additional set of elements, whose values ​​are derived from an encoding of the first set of image data 962 using a neural network architecture such as 522 in FIG. 5. These additional elements may represent a "face encoding" and vector portion 928 may represent a "lip encoding." The face encoding may remain static for an utterance, and the lip encoding may change during an utterance or may include multiple "frames" for an utterance. Although FIG. 9 shows an example using both lip feature extractor 924 and feature extraction component 922, in one example, feature extraction component 922 may be omitted. In the following example, a lip reading system for use in a vehicle may be used in a manner similar to speech processing device 500 of FIG. 5.

[0092] Exemplary Motor Vehicle 10A and 10B show an example where the vehicle described herein is an automobile. FIG. 10A shows a side view 1000 of an automobile 1005. The automobile 1005 includes a control unit 1010 for controlling the components of the automobile 1005. The components of the audio processing device 120 as shown in FIG. 1B (and other examples) may be integrated into this control unit 1010. In other cases, the components of the audio processing device 120 may be realized as separate units with an option of connection with the control unit 1010. The automobile 1005 further includes at least one image capture device 1015. For example, the at least one image capture device 1015 may include the image capture device 110 shown in FIG. 1A. In this example, the at least one image capture device 1015 may be communicatively coupled to the control unit 1010 and may be controlled by the control unit 1010. In another example, the at least one image capture device 1015 is in communication with the control unit 1010 and is remotely controlled. With the functionality described herein, the at least one image capture device 1015 may be used for video communications, such as voice over Internet Protocol calls with video data, environmental monitoring, driver alertness monitoring, etc. Figure 10A further shows at least one audio capture device in the form of a side-mounted microphone 1020. These may implement the audio capture device 116 shown in Figure 1A.

[0093] The image capture devices described herein may include one or more still or video cameras configured to capture frames of image data on command or at a predefined sampling rate. The image capture devices may provide coverage of both the front and rear of the vehicle interior. In one case, the predefined sampling rate may be less than the frame rate for full resolution video, e.g., a video stream may be captured at 30 frames / second, but the sampling rate of the image capture device may capture at this rate or at a lower rate, such as 1 frame / second. The image capture device may capture one or more frames of image data having one or more color channels (e.g., RGB or YUV as described above). In some cases, the frame rate and the frame size ... 30 frames / second, but the sampling rate of the image capture device may capture at this rate or at a lower rate, such as 1 frame / second. Aspects of the image capture device such as size and resolution, number of color channels, and sample format may be configurable. In some cases, frames of image data may be downsampled. For example, a video capture device that captures video at a "4K" resolution of 3840x2160 may be downsampled to 640x480 or less. Alternatively, for low-cost embedded devices, a low-resolution image capture device that captures frames of image data at 320x240 or less may be used. In some cases, even a cheap low-resolution image capture device may provide enough visual information to improve on the audio processing. As mentioned above, the image capture device may further include image pre-processing and / or filtering components (e.g., contrast adjustment, noise removal, color adjustment, cropping, etc.). In some cases, low-latency and / or high frame rate image cameras are available that meet the more stringent Automotive Safety Integrity Level (ASIL) for the ISO26262 automotive safety standard. Apart from their safety benefits, providing higher temporal information may improve lip-reading accuracy. It can be useful for recurrent neural networks for more accurate feature probability estimation.

[0094] FIG. 10B shows an overhead view 1030 of a car 1005. The car 1005 includes a front seat 1032 and a rear seat 1034 for holding passengers facing a front-mounted microphone for voice capture. The car 1005 includes a driver vision console 1036 with safety-critical display information. The driver vision console 1036 may include a portion of the dashboard 108 as shown in FIG. 1A. The car 1005 further includes a general console 1038 with navigation, entertainment and environmental control functions. The control unit 1010 may control the general console 1038 and may implement a local voice processing module and a wireless network communication module such as 120 in FIG. 1A. The wireless network communication module may transmit one or more of the image data, audio data and speaker feature vectors generated by the control unit 1010 to a remote server for processing. The automobile 1005 further includes a side mounted microphone 1020, a front overhead multi-microphone speech capture unit 1042, and a rear overhead multi-microphone speech capture unit 1044. The front and rear speech capture units 1042, 1044 provide additional audio capture devices for capturing speech audio, canceling noise, and identifying speaker location. In one case, the front and rear speech capture units 1042, 1044 may further include additional image capture devices for capturing image data characterizing each of the vehicle's occupants.

[0095] In the example of FIG. 10B, any one or more of the microphone and speech capture units 1020, 1042, and 1044 may provide audio data to an audio interface, such as 140 in FIG. 1B. The microphone or array of microphones may be configured to capture or record audio samples at a predefined sampling rate. In some cases, aspects of each audio capture device, such as sampling rate, bit resolution, number of channels, and sample format, may be configurable. The captured audio data may be Pulse Code Modulated. Any of the audio capture devices may further include audio pre-processing and / or filtering components (e.g., contrast adjustment, noise removal, etc.). Similarly, any one or more of the image capture devices may provide image data to an image interface, such as 150 in FIG. 1B, and may further include video pre-processing and / or filtering components (e.g., contrast adjustment, noise removal, etc.).

[0096] FIG. 11 shows an example of the interior of an automobile 1100 as seen from the front seat 1032. For example, FIG. 11 may include a view toward the windshield 104 of FIG. 1A. FIG. 11 shows a steering wheel 1106 (such as the steering wheel 106 in FIG. 1), a side microphone 1120 (such as one of the side microphones 1020 in FIGS. 10A and 10B), a rearview mirror 1142 (which may include a front overhead multi-microphone speech capture unit 1042), and a projection device 1130. The projection device 1130 may be used to project an image 1140 onto the windshield, for example, for use as an additional visual output device (e.g., in addition to the driver vision console 1036 and the general console 1038). In FIG. 11, the image 1140 includes directions. These may be directions projected following a voice command such as "give me directions to the mall mart." Other examples may use a simpler response system.

[0097] Local and Remote Audio Processing for Vehicles In some cases, the functions of the voice processing module may be distributed as described herein. For example, some functions may be computed locally in the automobile 1005 and some functions may be computed by a remote ("cloud") server device. In some cases, functions may be replicated on the automobile ("client") side and on the remote server device ("server") side. In these cases, if a connection to the remote server device is not available, processing may be performed by the local voice processing module, and if a connection to the remote server device is available, one or more of the audio data, image data, and speaker feature vectors may be sent to the remote server device for analyzing the captured utterance. The remote server device may have processing resources (e.g., a central processing unit or CPU, a graphical processing unit or GPU, and random access memory) and thus may provide an improvement over local performance when a connection is available. This may be traded off against latency in the processing pipeline (e.g., local processing is more responsive). In one case, the local voice processing module may provide a first output, which may be supplemented and / or enhanced by the results of the remote voice processing module.

[0098] In one case, a vehicle, such as, for example, automobile 1005, may be communicatively coupled to a remote server device via at least one network. The network may include one or more local and / or wide area networks that may be implemented using a variety of physical technologies (e.g., wired technologies such as Ethernet and / or wireless technologies such as Wi-Fi (IEEE 802.11 standard and cellular communication technologies). In some cases, the network may include a mix of one or more private and public networks, such as the Internet. The vehicle and the remote server device may communicate over the network using different technologies and communication pathways.

[0099] 3, in one case, vector generation by vector generator 372 may occur locally or remotely, but data store 374 resides locally within automobile 1005. In this case, static speaker feature vectors may be computed locally and / or remotely, but are stored locally within data store 374. Speaker feature vectors 325 may then be extracted from data store 374 within the automobile, rather than received from a remote server device. This may improve speech processing latency.

[0100] If the voice processing module is remote from the vehicle, the local voice processing device includes a transceiver for transmitting data derived from one or more of the audio data, the image data, and the speaker feature vector to the voice processing module and for receiving control data from the analysis of the utterance. In one case, the transceiver may include a wired or wireless physical interface and one or more communication protocols that provide a method for sending and / or receiving requests in a predefined format. In one case, the transceiver may include an application layer interface that operates on the Internet Protocol Suite. In this case, the application layer interface may be configured to receive communications directed toward a specific Internet Protocol address that identifies a remote server device, with routing based on a path name or web address being performed by one or more proxies and / or communication (e.g., "web") servers.

[0101] In some cases, linguistic features generated by the voice processing module may be mapped to a voice command and a set of data for the voice command (e.g., as described with reference to the utterance parser 436 in FIG. 4). In one case, the utterance data 442 may be used by the control unit 1010 of the automobile 1005 to execute the voice command. In one case, the utterance parser 436 may reside in a remote server device, and the utterance analysis may involve identifying an appropriate service to execute the voice command from the output of the voice processing module. For example, the utterance parser 436 may communicate an application programming interface (API) to the identified server. The vehicle may be configured to make a Directions Service API request, which includes the command identified from the output of the language model and any command data. For example, an utterance of "Where is the mall mart?" may result in a text output of "Where is the mall mart," which may be mapped to a Directions Service API request for vehicle mapping data having a desired location parameter of "mall mart" and the vehicle's current location derived from a positioning system, such as the Global Positioning System. A response may be extracted and communicated to the vehicle and displayed as shown in FIG. 11.

[0102] In one case, the remote utterance parser 436 communicates the response data to the control unit 1010 of the automobile 1005. This may include machine-readable data that is communicated to the user, for example, via a user interface or audio output. The response data may be processed and a response to the user may be output on one or more of the driver vision console 1036 and the general console 1038. Providing a response to the user may include displaying text and / or images on a display screen of one or more of the driver vision console 1036 and the general console 1038, or may include outputting sound via a text-to-speech module. In some cases, the response data may include audio data that may be processed in the control unit 1005 and used to generate an audio output, for example, via one or more speakers. The response may be spoken to the user via a speaker mounted within the interior of the automobile 1005.

[0103] Exemplary Embedded Computing System FIG. 12 illustrates an exemplary embedded computing system 1200 that may implement an audio processing device as described herein. A system similar to the embedded computing system 1200 may be used to implement the control unit 1010 in FIG. 10. The exemplary embedded computing system 1200 includes one or more computer processor (CPU) cores 1210 and zero or more graphics processor (GPU) cores 1220. The processors connect through board-level wiring 1230 to random access memory (RAM) devices 1240 for program code and data storage. The embedded computing system 1200 further includes a network interface 1250 to allow the processor to communicate with remote systems and specific vehicle control circuits 1260. By executing instructions stored in the RAM devices through the interface 1230, the CPU 1210 and / or GPU 1220 may perform functions as described herein. In some cases, constrained embedded An embedded computing device may have a similar general arrangement of components, but in some cases may have fewer computing resources and may not have a dedicated graphics processor 1220.

[0104] Exemplary Audio Processing Method FIG. 13 illustrates an example method 1300 for processing audio that may be performed to improve in-vehicle voice recognition. Method 1300 begins at block 1305 where audio data is received from an audio capture device. The audio capture device may be located in the vehicle. The audio data may feature vocalizations from a user. Block 1305 includes capturing data from one or more microphones, such as devices 1020, 1042, and 1044 in FIGS. 10A and 10B. In one case, block 1305 may include receiving the audio data via a local audio interface, and in another case, block 1305 may include receiving the audio data over a network, for example, at an audio interface that is remote from the vehicle.

[0105] At block 1310, image data is received from an image capture device. The image capture device may be located within the vehicle and may include, for example, image capture device 1015 in Figures 10A and 10B. In one case, block 1310 may include receiving the image data via a local image interface, and in another case, block 1310 may include receiving the image data over a network, for example, at an image interface that is remote from the vehicle.

[0106] In block 1315, a speaker feature vector is obtained based on the image data. This may include, for example, implementing any one of the speaker pre-processing modules 220, 320, 520, and 920. Block 1315 may be performed by a local processor of the automobile 1005 or a remote server device. In block 1320, the utterance is analyzed using a voice processing module. For example, this may include, for example, implementing any one of the voice processing modules 230, 330, 400, 530, and 930. Block 1320 includes a number of sub-blocks. These include, in sub-block 1322, providing the speaker feature vector and audio data as input to an acoustic model of the voice processing module. This may include operations similar to those described with reference to FIG. 4. In some cases, the acoustic model includes a neural network architecture. In sub-block 1324, phoneme data is predicted using at least a neural network architecture based on the speaker feature vector and the audio data. This may include using a neural network architecture that is trained to receive speaker feature vectors as inputs in addition to audio data. Because both the speaker feature vectors and audio data include numerical representations, they may be similarly processed by neural network architectures. In some cases, an existing CTC or hybrid acoustic model may be configured to receive a concatenation of the speaker feature vector and audio data, and trained using a training set that additionally includes image data (e.g., used to derive the speaker feature vector).

[0107] In some cases, block 1315 includes performing facial recognition on the image data to identify people in the vehicle. For example, this may be done as described with reference to facial recognition module 370 in FIG. 3. After this, user profile data about the people (e.g., in the vehicle) may be obtained based on the facial recognition. For example, the user profile data may be extracted from data store 374 using user identifier 376, as described with reference to FIG. 3. The speaker feature vector is then added to the user profile data. In one case, the speaker feature vector may be extracted from the user profile data as a static set of element values. In another case, the user profile data may indicate that the speaker feature vector is to be calculated using one or more of the audio data and image data received, for example, at blocks 1305 and 1310. In one case, block 1315 includes comparing a number of stored speaker feature vectors associated with the user profile data to a predefined threshold. For example, the user profile data may indicate how many previous voice queries have been made by a user identified using face recognition. In response to the number of stored speaker feature vectors being below the predefined threshold, the speaker feature vector may be calculated using one or more of the audio data and image data. In response to the number of stored speaker feature vectors being above the predefined threshold, a static speaker feature vector may be obtained, for example, a speaker feature vector stored in or accessible via the user profile data. In this case, the static speaker feature vector may be generated using a number of stored speaker feature vectors.

[0108] In one example, block 1315 may include processing the image data to generate one or more speaker feature vectors based on lip movements within the person's facial area. For example, a lip reading module such as lip feature extractor 924 or a suitably configured neural speaker pre-processing module 520 may be used. The output of the lip reading module may be used to feed one or more speaker feature vectors to a speech processing module and / or may be combined with other values ​​(e.g., i or x vectors) to generate a larger speaker feature vector.

[0109] In one example, block 1320 includes providing phoneme data to a language model of a voice processing module, predicting a transcript of the utterance using the language model, and determining control commands for the vehicle using the transcript. For example, block 1320 may include operations similar to those described with reference to FIG.

[0110] Voice analysis example 14 illustrates an example processing system 1400 including a non-transitory computer-readable storage medium 1410 storing instructions 1420. The instructions 1420, when executed by at least one processor 1430, cause the at least one processor to perform a series of operations. The operations in this example use approaches previously described for generating transcriptions of utterances. These operations may be performed, for example, in a vehicle as described above, or may extend the in-vehicle example to non-vehicle-based situations that may be implemented using, for example, a desktop, laptop, mobile or server computing device.

[0111] Via instructions 1432, the processor 1430 is configured to receive audio data from an audio capture device. This may include accessing a local memory containing the audio data and / or receiving a data stream or set of array values ​​over a network. The audio data may have a form as described with reference to other examples herein. Via instructions 1434, the processor 1430 is configured to receive a speaker feature vector. The speaker feature vector is obtained based on image data from an image capture device, the image data characterizing a facial area of ​​the user. For example, the speaker feature vector may be obtained using the approaches described with reference to any of Figures 2, 3, 5 and 9. The speaker feature vector may, for example, be calculated locally by the processor 1430 and retrieved from the local memory. The voice processing module 1430 may include any of the modules described with reference to any of Figures 2, 3, 4, 5, and 9. ...

[0112] FIG. 14 shows that the instructions 1436 can be broken down into a number of further instructions. Via instructions 1440, the processor 1430 is instructed to provide the speaker feature vector and the audio data as input to an acoustic model of the voice processing module. This can be accomplished in a manner similar to that described with reference to FIG. 4. In this example, the acoustic model includes a neural network architecture. Via instructions 1442, the processor 1430 is instructed to predict phoneme data based on the speaker feature vector and the audio data, using at least the neural network architecture. Via instructions 1444, the processor 1430 is instructed to provide the phoneme data to a language model of the voice processing module. This can be further performed in a manner similar to that shown in FIG. 4. Via instructions 1446, the processor 1430 is instructed to generate a transcript of the utterance using the language model. For example, the transcript can be generated as an output of the language model. In some cases, the transcript can be used by a control system, such as the control unit 1010 in the automobile 1005, to execute the voice command. In other cases, the transcript may include the output for a speech-to-text system. In the latter case, the image data may be extracted from a webcam or the like that is communicatively coupled to a computing device that includes the processor 1430. For mobile computing devices, the image data may be obtained from a front-facing image capture device.

[0113] In one example, the speaker feature vector received in accordance with instructions 1434 includes one or more of a speaker-dependent vector element (e.g., an i-vector or an x-vector component) generated based on the audio data, a speaker's lip movement-dependent vector element (e.g., as generated by a lip reading module) generated based on the image data, and a speaker's face-dependent vector element generated based on the image data. In one case, the processor 1430 may include part of a remote server device, and the audio data and speaker image vectors may be received from the autonomous vehicle, for example, as part of a distributed processing pipeline.

[0114] Example Implementation Certain examples are described that relate to speech processing, including automatic speech recognition. Certain examples relate to processing of certain spoken languages. Various examples work similarly for other languages ​​or language combinations. Certain examples improve the accuracy and robustness of speech processing by incorporating additional information derived from an image of the person making the utterance. This additional information can be used to improve a linguistic model. The linguistic model can include one or more of an acoustic model, a pronunciation model, and a language model.

[0115] Certain examples described herein may be implemented to address the unique challenges of performing automatic speech recognition in vehicles such as automobiles. In a combined example, image data from a camera may be used to determine lip reading features and recognize faces to enable i-vector and / or x-vector profiles to be constructed and selected. By implementing the approach as described herein, it may be possible to perform automatic speech recognition in the noisy, multi-channel environment of an automotive vehicle.

[0116] Certain examples described herein may include, for example, audio input or audio and The efficiency of speech processing may be increased by including one or more features derived from image data, such as lip position or movement, in a speaker feature vector that is provided as input to an acoustic model that also receives audio data as input (singular model), rather than having an acoustic model that only receives image data and a separate acoustic model for image data.

[0117] Certain methods and sets of operations may be performed by instructions stored on a non-transitory computer-readable medium. The non-transitory computer-readable medium stores code including instructions that, when executed by one or more computers, cause the computer to perform steps of the methods described herein. The non-transitory computer-readable medium may include one or more of a rotating magnetic disk, a rotating optical disk, a flash random access memory (RAM) chip, and other mechanically moving or solid-state storage media. Any type of computer-readable medium is suitable for storing code including instructions according to various examples.

[0118] Certain examples described herein may be implemented as so-called system-on-chip (SoC) devices. SoC devices may be used to control many embedded in-vehicle systems and implement the functionality described herein. In one case, one or more of the speaker pre-processing module and the voice processing module may be implemented as an SoC device. An SoC device may include one or more processors (e.g., CPU or GPU), random access memory (e.g., RAM, such as off-chip dynamic RAM or DRAM), and a variety of communication technologies, including Ethernet, WiFi, 3G, 4G long-term evolution (LTE), 5G, etc. , and network interfaces for wired or wireless connections, such as radio and other wireless interface standards. The SoC device may further include various I / O interface devices as required for different peripherals, for example, touch screen sensors, geolocation receivers, microphones, speakers, Bluetooth peripherals, and USB devices such as keyboards and mice. By executing instructions stored in the RAM device, the processor of the SoC device may perform the steps of the methods as described herein.

[0119] Although certain examples have been described herein, it is noted that different combinations of different components may be possible in different examples. Although salient features have been shown to better explain the examples, it will be apparent that certain features may be added, modified and / or omitted without modifying the functional aspects of these examples as described.

[0120] Various examples are methods that use either or a combination of human and machine behavior. The method examples are complete since most of the constituent steps occur anywhere in the world. Some examples are one or more non-transitory computer-readable media configured to store such instructions for the methods described herein. Examples can be realized on any machine that holds a non-transitory computer-readable medium that includes any of the necessary code. Some examples can be realized as physical devices such as semiconductor chips, hardware description language representations of the logical or functional behavior of such devices, and one or more non-transitory computer-readable media configured to store such hardware description language representations. Descriptions herein that describe principles, aspects, and embodiments encompass structural and functional equivalents thereof. Elements described herein as being coupled have an effective relationship that can be realized by a direct connection or indirectly with one or more other intervening elements.

Claims

1. an audio interface configured to receive audio data from an audio capture device of the vehicle; an image interface configured to receive image data from an image capture device to capture an image from the vehicle; a voice processing module configured to analyze human speech based on the audio data and the image data; a speaker pre-processing module configured to receive the image data and obtain a speaker feature vector based on the image data to predict phoneme data.

2. The in-vehicle device of claim 1 , wherein the voice processing module includes an acoustic model configured to process the audio data and predict the phoneme data used to analyze the utterance.

3. The in-vehicle device of claim 2 , wherein the acoustic model comprises a neural network architecture.

4. 4. The in-vehicle device of claim 2 or 3, wherein the acoustic model is configured to receive the speaker feature vector and the audio data as input and is trained to use the speaker feature vector and the audio data to predict the phoneme data.

5. The in-vehicle device according to claim 1, wherein the image data includes a face area of ​​a person inside the vehicle.

6. The speaker pre-processing module includes: performing facial recognition on the image data to identify the person within the vehicle; and extracting a speaker feature vector associated with the identified person.

7. The speaker pre-processing module includes:

10. An in-vehicle device according to claim 1 or 6, comprising a processor-implemented lip reading module, the lip reading module configured to generate one or more speaker feature vectors based on lip movement within the facial area of ​​the person.

8. The in-vehicle device of any one of claims 1 to 7, wherein the speaker pre-processing module includes a neural network architecture, the neural network architecture configured to receive data derived from one or more of the audio data and the image data and to predict the speaker feature vector.

9. The speaker pre-processing module includes: Calculating a speaker feature vector for a predefined number of utterances; 9. The in-vehicle device according to claim 1, configured to: calculate a static speaker feature vector based on a plurality of said speaker feature vectors for said predefined number of utterances.

10. The in-vehicle device of claim 9 , wherein the static speaker feature vector is stored in a memory of the vehicle.

11. a memory configured to store one or more user profiles; The speaker pre-processing module includes: performing facial recognition on the image data to identify a user profile in the memory associated with the person in the vehicle; calculating a speaker feature vector for the person; storing the speaker feature vector in the memory; An in-vehicle device according to any preceding claim, configured to: associate the stored speaker feature vector with the identified user profile.

12. The speaker pre-processing module includes: determining whether a number of stored speaker feature vectors associated with a given user profile is greater than a predefined threshold; in response to exceeding the predefined threshold, calculating a static speaker feature vector based on the number of stored speaker feature vectors; storing said static speaker feature vector in said memory; Associating the stored static speaker feature vector with the given user profile; and indicating that the static speaker feature vector is to be used for future utterance analysis in lieu of calculating the speaker feature vector for the person.

13. An in-vehicle apparatus according to any preceding claim, comprising an image capture device configured to capture electromagnetic radiation having infrared wavelengths, said image capture device configured to send said image data to said image interface.

14. the speaker pre-processing module is configured to process the image data to extract one or more portions of the image data; An in-vehicle device according to any one of claims 1 to 13, wherein the extracted one or more portions are used to obtain the speaker feature vector.

15. The on-vehicle device of any one of claims 1 to 14, wherein the processor is located within the vehicle.

16. the audio processing module is remote from the vehicle, and the apparatus comprises: An in-vehicle device according to any preceding claim, comprising a transceiver for transmitting data derived from the audio data and the image data to the voice processing module and for receiving control data from the analysis of the vocalisation.

17. 17. The in-vehicle device according to claim 1, wherein the acoustic model includes a hybrid acoustic model including a neural network architecture and a Gaussian mixture model, the Gaussian mixture model being configured to receive a vector of class probabilities output by the neural network architecture and to output phoneme data for analyzing the utterance.

18. The in-vehicle device according to any one of claims 1 to 17, wherein the acoustic model includes a connectionist temporal classification (CTC) model.

19. The audio processing module includes: An in-vehicle device according to any preceding claim, comprising a language model communicatively coupled to the acoustic model for receiving the phoneme data and generating a transcription representative of the utterance.

20. 20. The in-vehicle device of claim 19, wherein the language model is configured to use the speaker feature vector to generate the transcription representing the utterance.

21. The acoustic model is a database of acoustic model configurations; an acoustic model selector for selecting an acoustic model configuration from the database based on the speaker feature vector; and an acoustic model instance for processing the audio data, the acoustic model instance being instantiated based on the acoustic model configuration selected by the acoustic model selector, the acoustic model instance being configured to generate the phoneme data used to analyze the utterance.

22. The in-vehicle device according to any one of claims 1 to 21, wherein the speaker feature vector is one or more of an i-vector and an x-vector.

23. The speaker feature vector is a speaker-dependent first portion generated based on the audio data; and a second part dependent on the movement of the lips of the speaker, the second part being generated based on the image data.

24. 24. The in-vehicle device of claim 23, wherein the speaker feature vector includes a third portion that is generated based on the image data and that depends on a face of the speaker.

25. receiving audio data from an audio capture device located within the vehicle, the audio data characterized by vocalizations of persons within the vehicle; receiving image data from an image capture device to capture an image within the vehicle, the image data characterizing a facial area of ​​the person; obtaining a speaker feature vector based on the image data; and analyzing the utterance using a processor-implemented speech processing module; Analyzing the utterance includes: providing the speaker feature vector and the audio data as input to an acoustic model of the speech processing module, the acoustic model including a neural network architecture; and 20. A method for processing an utterance comprising predicting phoneme data based on said speaker feature vector and said audio data using at least said neural network architecture.

26. Obtaining a speaker feature vector includes the steps of: performing facial recognition on the image data to identify the person within the vehicle; obtaining user profile data about the person based on the facial recognition; and deriving the speaker feature vector in accordance with the user profile data.

27. A number of stored speaker feature vectors associated with the user profile data is predicted. comparing the detected value with a threshold value specified for the responsive to the number of stored speaker feature vectors falling below the predefined threshold, calculating the speaker feature vectors using one or more of the audio data and the image data; 27. The method of claim 26, comprising: in response to the number of stored speaker feature vectors being greater than the predefined threshold, obtaining a static speaker feature vector associated with the user profile data, the static speaker feature vector being generated using the number of stored speaker feature vectors.

28. Obtaining a speaker feature vector includes the steps of: A method according to any one of claims 25 to 27, comprising processing the image data to generate one or more speaker feature vectors based on lip movement in the facial area of ​​the person.

29. Analyzing the utterance includes: providing said phoneme data to a language model of said speech processing module; predicting a transcript of the utterance using the language model; and and using the transcript to determine control commands for the vehicle.

30. 16. A program code comprising instructions, which when executed by at least one processor, cause the at least one processor to: receiving audio data from an audio capture device; receiving a speaker feature vector derived based on image data from an image capture device, the image data characterizing a facial area of ​​the user; and analyzing the utterance using a speech processing module; Analyzing the utterance includes: providing the speaker feature vector and the audio data as input to an acoustic model of the speech processing module, the acoustic model including a neural network architecture; and predicting phoneme data based on the speaker feature vector and the audio data using at least the neural network architecture; providing said phoneme data to a language model of said speech processing module; and generating a transcript of the utterance using the language model.

31. The speaker feature vector is speaker dependent vector elements generated based on the audio data; A vector element dependent on the movement of the lips of the speaker, the vector element being generated based on the image data; and a vector element dependent on the speaker's face, the vector element being generated based on the image data.

32. 31. The program code of claim 30, wherein the audio data and the speaker image vector are received from a motor vehicle.

Citation Information

Patent Citations

  • Bathroom abnormality sensing system, bathroom abnormality sensing method, and bathroom abnormality sensing program

    JP2016170701A

  • Voice recognition control system

    JP2017090612A

  • Passenger state detection device, passenger state detection system, and passenger state detection method

    WO2019159364A1

  • Methods using recurrence quantification analysis to analyze and generate images

    US8422820B2