An optimization method, device and computer program product of a vehicle intelligent cockpit voice wake-up model

Abstract

This invention discloses an optimization method, device, and computer program product for a voice wake-up model in a vehicle's intelligent cockpit. The method includes: collecting multi-modal user behavior data, cockpit audio data, and environmental data according to voice false wake-up data collection trigger rules; parsing the collected data to obtain structured data related to voice false wake-ups; extracting data associated with voice false wake-ups from the structured data; classifying the data associated with voice false wake-ups to form a false wake-up training dataset; and importing the false wake-up training dataset into the voice wake-up model for iterative training to optimize the voice wake-up model. This invention not only helps reduce false wake-up events and significantly lowers the false wake-up rate, but also enables the voice wake-up model to better adapt to actual usage scenarios, reduce unnecessary interference, and provide users with a more targeted voice interaction experience, while enhancing the model's self-learning and continuous optimization capabilities.

Description

Technical Field

[0001] This invention relates to the field of intelligent connected vehicle technology, specifically to an optimization method, device, and computer program product for a vehicle intelligent cockpit voice wake-up model. Background Technology

[0002] In the current field of voice wake-up technology for smart cockpits, the mainstream approach mainly focuses on audio frequency matching analysis, that is, activating the wake-up function by comparing the audio signal emitted by the user with the audio frequency characteristics of a preset wake-up word. However, this technical approach has revealed significant limitations in practical applications, as it ignores more complex and crucial auxiliary judgment conditions, such as understanding the context and verifying the authenticity of the sound source.

[0003] Specifically, current technical solutions typically follow this process: First, effective audio detection is performed using VW-VAD (VoiceWake-up Voice Activity Detection) technology to eliminate silent segments and non-wake-up related audio, thereby optimizing computational resource allocation. Next, the feature extraction stage is entered, using the FBank (Filter Bank) algorithm to extract frequency band or frequency point sampling features of the audio, providing input data for the acoustic model. Then, the acoustic model is responsible for classifying the pronunciation state of each 25-millisecond audio frame. Afterward, through the decoding process, a decoding path is constructed based on the wake word, and the matching degree between the audio and the wake word is evaluated. Finally, a confidence model is introduced for secondary verification of the wake-up results.

[0004] Although the above solutions have achieved voice wake-up functionality to some extent, they still have the following shortcomings:

[0005] Limitations of training data: The training of wake-up models relies heavily on a large amount of test data in a laboratory environment. This data deviates significantly from the complex and variable conditions in real-world driving scenarios, resulting in insufficient representativeness and generalization ability of the training data.

[0006] A singular perspective on problem-solving: When faced with the complex problem of false wake-up, existing technologies often seek solutions based on a single factor, ignoring the combined influence of multiple possible factors, which limits the improvement of false wake-up suppression.

[0007] Lack of in-depth data analysis: Few technical solutions conduct in-depth data analysis on the specific reasons for false voice wake-up, failing to accurately identify the root cause of the problem, thus making it difficult to formulate targeted optimization strategies to fundamentally solve the false wake-up problem. Summary of the Invention

[0008] The technical problem to be solved by the embodiments of the present invention is to provide an optimization method, device and computer program product for a voice wake-up model of a vehicle intelligent cockpit, so as to effectively reduce the false wake-up rate of voice and improve the accuracy of the voice recognition system and the user experience.

[0009] To address the aforementioned technical problems, this invention provides an optimization method for a vehicle intelligent cockpit voice wake-up model, comprising the following steps:

[0010] Based on the voice false wake-up data collection trigger rules, collect multi-modal user behavior data, in-cabin audio data, and environmental data;

[0011] Analyze the collected data to obtain structured data related to voice-activated false wake-up;

[0012] Extract data associated with voice-activated false wake-up from the structured data;

[0013] The data associated with voice-activated false wake-ups are classified to form a false wake-up training dataset;

[0014] The false wake-up training dataset is imported into the voice wake-up model for iterative training to optimize the voice wake-up model.

[0015] Preferably, the step of collecting user behavior multimodal data, in-cabin audio data, and environmental data according to the voice false wake-up data collection triggering rules specifically includes:

[0016] It detects and responds to voice wake-up signals and recognizes the wake-up voice to confirm whether it is a valid wake-up command;

[0017] The recognized wake-up voice is compared with a homonym database to determine whether the comparison was successful.

[0018] If the comparison is successful, the monitoring of the next round of voice interaction will continue; if the comparison is unsuccessful, it will be judged as a suspected false wake-up, and the collection of user behavior multimodal data, cabin audio data and environmental data will be triggered.

[0019] If no valid wake-up command is recognized in the next round of voice interaction, it is determined to be a suspected false wake-up, and the collection of user behavior multi-modal data, cabin audio data and environmental data is triggered.

[0020] Preferably, the structured data suitable for machine learning specifically includes user feature data and wake-up environment data.

[0021] Preferably, the step of extracting data associated with voice-induced false wake-up from the structured data specifically includes:

[0022] Features related to voice-induced false wake-up are selected and extracted from the structured data;

[0023] The association rule mining algorithm was used to analyze the relationship between the extracted features and false voice wake-up.

[0024] Based on the aforementioned correlation, data directly related to voice-activated false wake-up was extracted.

[0025] Preferably, the step of classifying the data associated with voice-induced false wake-up to form a false wake-up training dataset specifically includes:

[0026] By analyzing the features related to voice-induced false wake-up in the structured data, the cause of the false wake-up can be determined;

[0027] Based on the cause of false wake-up, a false wake-up classification algorithm is used to classify the data related to voice false wake-up mined from structured data into non-false wake-up, accidental touch, echo self-excitation, human voice wake-up, external sound source, and non-effective sound wake-up.

[0028] Preferably, the method further includes:

[0029] A second-level confidence algorithm module is set in the voice wake-up model to calculate the second-level confidence of the current voice wake-up signal when the wake-up score of the current voice wake-up signal exceeds the preset wake-up threshold.

[0030] The secondary confidence level is used to compare with a preset secondary confidence threshold. If the secondary confidence level exceeds the preset secondary confidence threshold, the wake-up is determined to be successful.

[0031] Preferably, the step of calculating the second-order confidence level of the current voice wake-up signal using a two-order confidence algorithm specifically includes:

[0032] Using a trained computational model derived from features extracted from the false wake-up training dataset that reflect audio content, environmental conditions, and user characteristics, a confidence score is calculated to determine whether the current voice wake-up signal is a valid wake-up command.

[0033] The present invention also provides an optimization device for a voice wake-up model of a vehicle intelligent cockpit, comprising:

[0034] The data acquisition module is used to collect multi-modal user behavior data, in-cabin audio data, and environmental data according to the data acquisition trigger rules for voice-activated false wake-up.

[0035] The parsing module is used to parse the collected data and obtain structured data related to voice-induced false wake-up.

[0036] An extraction module is used to extract data related to voice false wake-up from the structured data;

[0037] The classification module is used to classify the data associated with voice-induced false wake-up and form a false wake-up training dataset;

[0038] An optimization module is used to import the false wake-up training dataset into the voice wake-up model for iterative training in order to optimize the voice wake-up model.

[0039] The present invention also provides an optimization device for a voice wake-up model of a vehicle intelligent cockpit, comprising:

[0040] One or more processors;

[0041] Memory;

[0042] One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, and the one or more applications are configured to perform the optimized method of the vehicle smart cockpit voice wake-up model.

[0043] The present invention also provides a computer program product, including computer instructions that instruct a computer device to perform an operation corresponding to the method.

[0044] Implementing this invention offers the following advantages: By comprehensively collecting multi-modal data including user behavior, in-cabin audio, and environmental data, this invention provides a rich data foundation for model optimization. Using false wake-up data as negative examples for iterative model training not only optimizes the model algorithm but also significantly reduces the false wake-up rate and improves wake-up accuracy. This invention enables the voice wake-up model to better adapt to real-world usage scenarios, reducing unnecessary interference and providing users with a more targeted voice interaction experience. It also enhances the model's self-learning and continuous optimization capabilities, providing strong support for the intelligent development of vehicle smart cockpits. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating an optimization method for a voice wake-up model in a vehicle intelligent cockpit according to an embodiment of the present invention.

[0047] Figure 2 This is a schematic diagram of the data acquisition process in an embodiment of the present invention.

[0048] Figure 3This is a schematic diagram of the second-level confidence level judgment process in an embodiment of the present invention. Detailed Implementation

[0049] The following description of the embodiments is taken with reference to the accompanying drawings, which illustrate specific embodiments in which the invention can be implemented.

[0050] Please refer to Figure 1 As shown, Embodiment 1 of the present invention provides an optimization method for a voice wake-up model of a vehicle intelligent cockpit, comprising the following steps:

[0051] Based on the voice false wake-up data collection trigger rules, collect multi-modal user behavior data, in-cabin audio data, and environmental data;

[0052] Analyze the collected data to obtain structured data suitable for machine learning;

[0053] Extract data associated with voice-activated false wake-up from the structured data;

[0054] The data associated with voice-activated false wake-ups are classified to form a false wake-up training dataset;

[0055] The false wake-up training dataset is imported into the voice wake-up model for iterative training to optimize the voice wake-up model.

[0056] As demonstrated by the above steps, this invention provides a rich data foundation for model optimization by comprehensively collecting multi-modal data, including user behavior, in-cabin audio, and environmental data. By using false wake-up data as negative examples for iterative model training, it not only optimizes the model algorithm but also significantly reduces the false wake-up rate and improves wake-up accuracy. This invention enables the voice wake-up model to better adapt to real-world usage scenarios, reduce unnecessary interference, and provide users with a more targeted voice interaction experience. It also enhances the model's self-learning and continuous optimization capabilities, providing strong support for the intelligent development of vehicle smart cockpits.

[0057] Specifically, please combine Figure 2 As shown in the embodiment of the present invention, the specific process for collecting multi-modal user behavior data, in-cabin audio data, and environmental data according to the voice false wake-up data collection triggering rules is as follows:

[0058] First, the voice wake-up signal is detected and responded to. Then, the wake-up voice is recognized to confirm whether it is a valid wake-up command.

[0059] The recognition result is then compared with a homophone database to determine if the comparison was successful. If the comparison is successful, the system continues monitoring the next round of voice interaction; if the comparison fails, it is determined to be a suspected false wake-up and the data collection mechanism is triggered.

[0060] It is understood that, in this embodiment of the invention, the purpose of comparing the recognized speech with a homophone dictionary is primarily to verify whether the recognized wake-up speech matches a preset wake-up word or command word. If the comparison is successful, it means that the user's speech matches a valid wake-up word or command word in the system. This typically indicates that the user has successfully interacted with the system via voice commands, and the system should execute the corresponding operation based on the recognized command. If the comparison fails, i.e., the wake-up text similarity is below a preset threshold (the similarity between the recognized wake-up text and the valid wake-up word is too low), it may mean that the user's speech does not match any valid wake-up word or command word in the system. This could be because the user did not utter the correct wake-up word, or the speech recognition system failed to accurately recognize the user's speech. Therefore, it will be judged as a false wake-up, and data collection will be triggered to gather relevant data for subsequent analysis and model optimization.

[0061] If no valid wake-up command is recognized in the next round of voice interaction, it is also judged as a suspected false wake-up, and the data collection mechanism is triggered. Of course, if the next round of voice interaction is still valid, it will maintain normal operation and continue to process user commands. That is to say, the voice false wake-up data collection triggering rules in this embodiment of the invention specifically include: the similarity of the wake-up text is lower than a preset threshold or the user does not have a valid wake-up command in the next round of voice interaction.

[0062] Table 1 below illustrates how to collect key data through data tracking strategies to address the false wake-up problem in smart cockpit voice wake-up. Table 1 lists several key influencing factors and the data collection strategies for these factors.

[0063] Table 1 Key factors affecting voice wake-up and their corresponding data tracking strategies

[0064]

[0065] Background noise is affected by both the cabin interior (e.g., audio playback, air conditioning, windows) and the external environment. Therefore, data needs to be collected from multiple perspectives, including the cabin's internal equipment (e.g., audio playback, air conditioning, windows) and the external environment. This data helps to identify and quantify the impact of background noise on speech recognition accuracy.

[0066] User characteristics include voice features (such as accent, speech rate, and tone) and behavioral characteristics. By collecting and analyzing on-site audio data, and combining it with anonymized data obtained through DMS technology, user voice features can be analyzed to help determine false triggering of the wake-up system and improve user experience.

[0067] Contextualized speech recognition refers to optimizing the accuracy of speech recognition based on the specific context in which the user is situated. By collecting data from the next round of interaction, the system can better understand the user's intentions and needs, thereby improving the accuracy of recognition and the appropriateness of the response.

[0068] Wake-up behavior characteristics involve the user's behavioral patterns in waking up the intelligent cockpit system. By collecting data from the wake-up site, it is possible to analyze the user's wake-up habits, the frequency of wake-up words used, etc., thereby optimizing the wake-up algorithm and improving the system's response speed.

[0069] As can be seen from the above, the data acquisition mechanism of this invention specifically includes the following aspects:

[0070] Multimodal user behavior data collection: Capture and record various user behavior data, such as key presses and gestures, to analyze the correlation between user behavior patterns and false wake-ups.

[0071] In-cabin audio data acquisition: This includes acquiring current audio data to record the voice characteristics during accidental wake-up.

[0072] In-cabin environmental data collection: This includes collecting environmental data inside the cabin, such as noise and temperature, to assess the impact of environmental factors on the accuracy of speech recognition.

[0073] Through the above process, the present invention can distinguish between valid user commands and invalid voice inputs, respond quickly when a suspected false wake-up is detected, and comprehensively collect relevant data to provide data support for subsequent reporting to the cloud platform for in-depth analysis, thereby continuously optimizing the accuracy of voice wake-up and user experience.

[0074] The collected data can be uploaded to a cloud platform for parsing and processing to obtain structured data suitable for machine learning. The specific process is as follows:

[0075] (1) Audio data sampling

[0076] Since the acquired audio data is a continuous analog signal, it cannot be processed directly by a computer. Therefore, an appropriate sampling rate (such as 16kHz, 44.1kHz, etc.) is used to sample the audio signal, converting the continuous audio waveform into a series of discrete numerical points.

[0077] (2) Audio data format conversion

[0078] The audio data needs to be converted to a suitable format for processing. The original audio files may be stored in different formats such as WAV, MP3, and AAC. For standardized processing, these data need to be converted to a common format, such as PCM (Pulse Code Modulation). During the conversion process, parameters such as the sampling rate and bit depth of the audio data must remain consistent to ensure the accuracy of subsequent processing.

[0079] (3) Audio signal preprocessing

[0080] Audio signals may contain noise, distortion, and other interference factors, requiring preprocessing to improve the quality of the audio signal and enhance the accuracy of subsequent feature extraction.

[0081] Preprocessing operations include:

[0082] Noise reduction: Using noise reduction algorithms (such as spectral subtraction, Kalman filtering, etc.) to reduce background noise and other interference, improving the clarity of the speech signal; Frequency response adjustment: Adjusting the frequency characteristics of the audio signal to adapt to specific processing needs.

[0083] Frequency response adjustment: Adjusting the frequency response of an audio signal using filters (such as high-pass filters, low-pass filters, band-pass filters, etc.) to suit specific processing needs.

[0084] Other preprocessing operations, such as audio enhancement and volume normalization, are used to further improve the quality of the audio signal.

[0085] (4) Feature extraction

[0086] Extracting features from the preprocessed audio signal that are helpful for subsequent processing or analysis, such as speech rate, pitch, and voiceprint. Specifically, this includes the following aspects:

[0087] Speech rate extraction: Measure the rate of speech, that is, the number of syllables spoken per unit of time.

[0088] Pitch extraction: Pitch detection algorithms (such as autocorrelation function method, Fourier transform method, etc.) are used to extract pitch information from audio signals.

[0089] Voiceprint extraction: Extracting voiceprint features from audio signals using voiceprint recognition techniques (such as Mel-frequency cepstral coefficients (MFCC), linear predictive coding (LPC)).

[0090] Other feature extraction methods include noise type and background music, to more comprehensively describe the characteristics of the audio signal.

[0091] (5) Feature data processing

[0092] The extracted features are further processed to obtain structured data suitable for machine learning. Specific operations include, but are not limited to:

[0093] Data cleaning: removing outliers and noise from feature data to improve data accuracy.

[0094] Normalization: Scaling feature data to a uniform range to reduce scale differences between different features.

[0095] Feature selection: Select the features that are most helpful in distinguishing different voice wake-up events.

[0096] Feature fusion: Fusing multiple features to improve the accuracy of voice wake-up.

[0097] Data encoding: Converting feature data into a format suitable for machine learning models to process.

[0098] Finally, the processed feature data is stored in a structured format in a database or data file for subsequent analysis and application. This structured data includes:

[0099] User characteristic data, such as accent, speech rate, and voiceprint, is used for personalized speech recognition and user verification.

[0100] Wake-up environment data: such as noise levels, which are used to understand the impact of the wake-up environment on speech recognition accuracy.

[0101] The structured data obtained through the above steps is not all directly related to voice-induced false wake-up. This embodiment of the invention further requires mining this structured data to obtain data directly related to voice-induced false wake-up. Specifically, this involves extracting data associated with voice-induced false wake-up by applying algorithms and statistical methods. Specifically, features potentially related to voice-induced false wake-up are selected from the structured data, such as vehicle speed, acceleration, whether the driver is speaking, the spectral characteristics of audio signals in the cabin, and environmental noise levels. Then, specific algorithms and statistical methods are applied to extract these features from the raw data. For example, signal processing techniques can be used to extract the spectral characteristics of audio signals, or machine learning algorithms can be used to recognize the driver's voice commands. Algorithms such as association rule mining are then used to analyze the correlation between the extracted features and voice-induced false wake-up to help identify which factors are most likely to cause false wake-up events. Finally, data associated with voice-induced false wake-up is mined.

[0102] After obtaining the data associated with voice-induced false wake-up, this embodiment of the invention employs a false wake-up classification algorithm. By analyzing features related to voice-induced false wake-up, such as audio content, sound source, and user behavior, the specific cause of the false wake-up is determined. This classifies the data related to voice-induced false wake-up mined from structured data into six main types: non-false wake-up, accidental touch, echo self-oscillation, human voice wake-up, external sound source, and ineffective sound wake-up.

[0103] Non-false wake-up: refers to a situation where a wake-up is actually valid but is incorrectly identified as a false wake-up;

[0104] Accidental touch: refers to a false wake-up caused by an unintentional touch or operation by the user;

[0105] Echo self-excitation: refers to false wake-up caused by sound reflection or echo in the cockpit;

[0106] Voice wake-up: refers to false wake-up caused by the voice of a non-specific user or other voice interference;

[0107] External sound sources: refers to false wake-up caused by sound sources such as music, broadcasts, and traffic noise from outside the cockpit;

[0108] Ineffective wake-up call: refers to false wake-up caused by audio that does not match the wake word or is of low quality.

[0109] The data for human voice wake-up type is further classified using a false wake-up word classification algorithm into similar words, similar command words, and irrelevant words. Finally, false wake-up training data is generated.

[0110] In this embodiment of the invention, generated false wake-up data is used as negative example data to input into the voice wake-up model for iterative training, thereby improving the accuracy and robustness of the voice wake-up model. It should be noted that, in this embodiment of the invention, negative example data refers to voice data that is incorrectly identified as a wake-up command by the voice wake-up model, but which should not actually trigger a wake-up. Negative example data contradicts the expected behavior of the voice wake-up model (i.e., responding only to valid wake-up commands), indicating that the voice wake-up model has defects or deficiencies in recognizing and processing voice wake-up requests.

[0111] Specifically, negative example data includes various scenarios of false wake-up by the model, such as wake-ups caused by environmental noise, non-specific human voices, echolocation, accidental touches, or other ineffective sound sources. These false wake-up events should not occur, but due to limitations in the model algorithm or insufficient training data, they are incorrectly identified as valid wake-up commands. By collecting these false wake-up events as negative example data and importing them into the voice wake-up model for iterative training, the model can learn under which circumstances wake-up should not be triggered, thereby optimizing its algorithm, improving wake-up accuracy, and reducing the false wake-up rate. Negative example data thus becomes an important basis for model optimization, helping the model identify and correct its previous erroneous judgments, enabling it to respond more accurately to valid wake-up commands in future use, while ignoring interfering factors that may lead to false wake-ups.

[0112] As a further improvement to the embodiments of the present invention, the embodiments of the present invention also add a two-level confidence confirmation process, that is, by adding an additional confidence judgment step, false wake-up events are reduced and the accuracy of voice wake-up is improved.

[0113] The second-order confidence level algorithm is a machine learning and multi-dimensional data analysis-based algorithm. It learns from the aforementioned counterexample data and combines it with various relevant data to calculate the second-order confidence threshold, thereby achieving a more accurate judgment of wake-up requests. The following is the specific process for calculating the second-order confidence threshold:

[0114] After preprocessing the collected counterexample data (including audio signal noise reduction, feature extraction, text transcription, etc.), features that reflect audio content, environmental conditions, user characteristics, and are associated with false wake-up events are extracted from it.

[0115] Machine learning algorithms (such as support vector machines, neural networks, etc.) are used to train the extracted features to build a model that can distinguish between valid wake-ups and false wake-ups;

[0116] Based on the trained model, predictions are made on the data in the validation set or test set, and the confidence score of each sample is calculated, which is the probability that the model considers the sample to be a valid wake-up.

[0117] Based on the relationship between confidence score and false wake-up rate, a second-level confidence threshold is calculated. The second-level confidence threshold can identify as many valid wake-ups as possible while ensuring a low false wake-up rate.

[0118] When applying the obtained second-level confidence threshold to wake-up judgment, the voice wake-up signal is calculated using a second-level confidence algorithm to obtain the corresponding second-level confidence score. This score is then compared with the second-level confidence threshold. If the score exceeds the threshold, it is considered a valid wake-up, triggering the corresponding response; if the score is below the threshold, it is considered a false wake-up or an invalid request, and no response is triggered. Please refer to [reference needed]. Figure 3 As shown below, the wake-up process is described in detail:

[0119] First, the current voice wake-up signal is collected, and then it is processed and recognized locally to determine whether it is a wake-up word.

[0120] The wake-up score is calculated based on the preliminary recognition results. The wake-up score is an evaluation of the degree of matching between the voice input and the wake-up word based on the local algorithm, taking into account various factors such as the clarity of the voice, background noise, and the speaker's voiceprint characteristics.

[0121] Determine if the wake-up score exceeds the preset wake-up threshold: Compare the wake-up score with the preset wake-up threshold. If the wake-up score exceeds the threshold, proceed to the next step; if the score does not exceed the threshold, it is determined as non-wake-up, the process ends, and the system does not respond.

[0122] Secondary confidence level determination: The current voice wake-up signal is calculated using a secondary confidence level algorithm to obtain the corresponding secondary confidence level; then it is determined whether the secondary confidence level score exceeds the secondary confidence level threshold. If the secondary confidence level score exceeds the secondary confidence level threshold, the wake-up is determined to be successful; if the secondary confidence level score does not exceed the secondary confidence level threshold, it is determined to be a false wake-up, the process ends, and the system does not respond.

[0123] This invention employs a two-stage judgment process. It initially filters potential wake-up attempts locally, then performs more precise confirmation in the cloud. This reduces false wake-ups caused by a single judgment criterion and improves the accuracy of wake-up event identification. In other words, it maintains a low false wake-up rate while enhancing the ability to respond to the user's true intent, thereby improving the user experience.

[0124] It should also be noted that the optimization method of the vehicle intelligent cockpit voice wake-up model in this embodiment of the invention can be continuously iterated through vehicle-cloud communication after the vehicle is sold.

[0125] Corresponding to the optimization method for the vehicle intelligent cockpit voice wake-up model described in Embodiment 1 of the present invention, Embodiment 2 of the present invention also provides an optimization device for the vehicle intelligent cockpit voice wake-up model, comprising:

[0126] The data acquisition module is used to collect multi-modal user behavior data, in-cabin audio data, and environmental data according to the data acquisition trigger rules for voice-activated false wake-up.

[0127] The parsing module is used to parse the collected data and obtain structured data related to voice-induced false wake-up.

[0128] An extraction module is used to extract data related to voice false wake-up from the structured data;

[0129] The classification module is used to classify the data associated with voice-induced false wake-up and form a false wake-up training dataset;

[0130] An optimization module is used to import the false wake-up training dataset into the voice wake-up model for iterative training in order to optimize the voice wake-up model.

[0131] Corresponding to the optimization method for the vehicle intelligent cockpit voice wake-up model described in Embodiment 1 of the present invention, Embodiment 3 of the present invention also provides an optimization device for the vehicle intelligent cockpit voice wake-up model, comprising:

[0132] One or more processors;

[0133] Memory;

[0134] One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, and the one or more applications are configured to perform the voice wake-up method of the automotive smart cockpit.

[0135] Corresponding to the optimization method of the vehicle intelligent cockpit voice wake-up model described in Embodiment 1 of the present invention, Embodiment 4 of the present invention also provides a computer program product, including computer instructions, wherein the computer instructions instruct a computer device to perform the operation corresponding to the method.

[0136] Preferably, the processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or the processor can be any conventional processor. The processor is the control center of the device, connecting various parts of the device through various interfaces and lines.

[0137] The memory mainly includes a program storage area and a data storage area. The program storage area can store the operating system, applications required for at least one function, etc., while the data storage area can store related data, etc. Furthermore, the memory can be a high-speed random access memory, or a non-volatile memory, such as a plug-in hard drive, a SmartMedia Card (SMC), a Secure Digital (SD) card, and a Flash Card, or other volatile solid-state storage devices.

[0138] It should be noted that the above-mentioned devices may include, but are not limited to, processors and memory, as will be understood by those skilled in the art.

[0139] For the working principle and process of the above embodiments, please refer to the description of Embodiment 1 of the present invention, which will not be repeated here.

[0140] As explained above, compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a rich data foundation for model optimization by comprehensively collecting multi-modal data such as user behavior, in-cabin audio, and environmental data; by using false wake-up data as negative examples to import into the model for iterative training, it not only optimizes the model algorithm but also significantly reduces the false wake-up rate and improves wake-up accuracy. This invention enables the voice wake-up model to better adapt to actual usage scenarios, reduce unnecessary interference, and bring users a more targeted voice interaction experience. It also enhances the model's self-learning and continuous optimization capabilities, providing strong support for the intelligent development of vehicle smart cockpits.

[0141] The above description is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. An optimization method for a voice wake-up model in a vehicle's intelligent cockpit, characterized in that, Includes the following steps: Based on the voice false wake-up data collection trigger rules, collect multi-modal user behavior data, in-cabin audio data, and environmental data; Analyze the collected data to obtain structured data containing information related to voice-activated false wake-up; Extract data directly related to voice-induced false wake-up from the structured data; Based on the cause of false wake-up, the data directly related to voice false wake-up is classified to form a false wake-up training dataset; The false wake-up training dataset is imported into the voice wake-up model for iterative training to optimize the voice wake-up model.

2. The method according to claim 1, characterized in that, The process of collecting user behavior multimodal data, in-cabin audio data, and environmental data based on the voice false wake-up data collection trigger rules specifically includes: It detects and responds to voice wake-up signals and recognizes the wake-up voice to confirm whether it is a valid wake-up command; The recognized wake-up voice is compared with a homonym database to determine whether the comparison was successful. If the comparison is successful, the monitoring of the next round of voice interaction will continue; if the comparison is unsuccessful, it will be judged as a suspected false wake-up, and the collection of user behavior multimodal data, cabin audio data and environmental data will be triggered. If no valid wake-up command is recognized in the next round of voice interaction, it is determined to be a suspected false wake-up, and the collection of user behavior multi-modal data, cabin audio data and environmental data is triggered.

3. The method according to claim 1, characterized in that, The structured data specifically includes user characteristic data and wake-up environment data.

4. The method according to claim 1, characterized in that, The extraction of data directly related to voice-induced false wake-up from the structured data specifically includes: Features related to voice false wake-up are selected and extracted from the structured data; The association rule mining algorithm was used to analyze the relationship between the extracted features and false voice wake-up. Based on the aforementioned correlation, data directly related to voice-activated false wake-up was extracted.

5. The method according to claim 4, characterized in that, The process of classifying data related to voice-based false wake-ups based on the cause of the false wake-up, forming a false wake-up training dataset, specifically includes: By analyzing the features related to voice-induced false wake-up in the structured data, the cause of the false wake-up can be determined; Based on the cause of false wake-up, a false wake-up classification algorithm is used to classify the data related to voice false wake-up mined from structured data into non-false wake-up, accidental touch, echo self-excitation, human voice wake-up, external sound source, and non-effective sound wake-up.

6. The method according to claim 1, characterized in that, Also includes: A second-level confidence algorithm module is set in the voice wake-up model to calculate the second-level confidence of the current voice wake-up signal when the wake-up score of the current voice wake-up signal exceeds the preset wake-up threshold. The secondary confidence level is used to compare with a preset secondary confidence threshold. If the secondary confidence level exceeds the preset secondary confidence threshold, the wake-up is determined to be successful.

7. The method according to claim 6, characterized in that, The calculation of the second-order confidence level of the current voice wake-up signal using the second-order confidence algorithm specifically includes: Using a trained computational model derived from features extracted from the false wake-up training dataset that reflect audio content, environmental conditions, and user characteristics, a confidence score is calculated to determine whether the current voice wake-up signal is a valid wake-up command.

8. An optimization device for a voice wake-up model in a vehicle's intelligent cockpit, characterized in that, include: The data acquisition module is used to collect multi-modal user behavior data, in-cabin audio data, and environmental data according to the data acquisition trigger rules for voice-activated false wake-up. The parsing module is used to parse the collected data and obtain structured data containing information related to voice-induced false wake-up. The extraction module is used to extract data directly related to voice false wake-up from the structured data; The classification module is used to classify the data associated with voice-based false wake-up based on the cause of the false wake-up, forming a false wake-up training dataset; An optimization module is used to import the false wake-up training dataset into the voice wake-up model for iterative training in order to optimize the voice wake-up model.

9. An optimization device for a voice wake-up model of a vehicle intelligent cockpit, characterized in that, include: One or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, and the one or more applications are configured to perform the optimization method for the vehicle intelligent cockpit voice wake-up model as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, Includes computer instructions that instruct a computer device to perform an operation corresponding to the method as described in any one of claims 1 to 7.

Images