Sound classification and localization using neural networks

Abstract

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for sound classification and localization using a neural network. One of the methods includes obtaining at least two sound signals captured by at least two audio sensors at two different locations at a first time point; generating an input characterizing the at least two sound signals; and processing the input characterizing the at least two sound signals using a neural network to generate a prediction at the first time point, wherein the prediction includes: (a) a respective score for each of a plurality of sound event classes indicating a predicted class for a first sound event detected in the at least two sound signals, and (b) a predicted location of a first sound source that emitted the first sound event.

Description

SOUND CLASSIFICATION AND LOCALIZATIONUSING NEURAL NETWORKSBACKGROUND

[0001] Sound event classification involves identifying what a sound signal represents and tagging sound signals into different classes of sound events, such as dog barking, human speech, siren, music, etc. Sound localization involves determining the location of a sound source using two or more audio sensors, such as multiple microphones.

[0002] Some existing techniques can perform sound localization using a first model for sound localization and / or sound separation, followed by sound event classification using a second model for sound event classification. For example, an existing system can first perform sound localization on a sound signal to determine a location of a sound source. The existing system can perform sound separation to isolate individual sounds from the sound source. Finally, the existing system can perform sound classification to determine a class for a sound event emitted from the sound source. Thus, the overall sound signal processing may not be accurate, may consume lots of computation resources, may take a long time to obtain results, or a combination of these.

[0003] Some existing techniques can only generate a single sound classification result and cannot accurately perform multiple classifications on a sound signal that captures multiple sound events. For example, if a sound signal captures a dog barking sound event concurrently with a human speech sound event, the existing techniques can only generate a single predicted sound event for the sound signal.

[0004] Some existing techniques cannot localize a sound signal emitted from multiple sound sources at different locations. For example, some existing techniques can only determine the location of the sound source that emits the loudest sound, e.g., after some frequency band filtering.SUMMARY

[0005] This specification describes systems and techniques for jointly performing both sound event classification and sound localization on an input sound signal using a combination of digital signal processing (DSP) techniques and neural network models.

[0006] In general, one innovative aspect of the subject matter described in this specification can be embodied in methods that include the actions of obtaining at least tw o sound signals captured by at least two audio sensors at two different locations at a first time point;generating an input characterizing the at least two sound signals; and processing the input characterizing the at least two sound signals using a neural network to generate a prediction at the first time point, wherein the prediction includes: (a) a respective score for each of a plurality of sound event classes indicating a predicted class for a first sound event detected in the at least two sound signals, and (b) a predicted location of a first sound source that emitted the first sound event. Other embodiments of this aspect include corresponding computer systems, apparatus, and computer programs recorded on one or more computer storage devices, each configured to perform the actions of the methods. For a system of one or more computers to be configured to perform particular operations or actions means that the system has installed on it software, firmware, hardware, or a combination of them that in operation cause the system to perform the operations or actions. For one or more computer programs to be configured to perform particular operations or actions means that the one or more programs include instructions that, when executed by data processing apparatus, cause the apparatus to perform the operations or actions.

[0007] The foregoing and other embodiments can each optionally include one or more of the following features, alone or in combination. In particular, one embodiment includes all the following features in combination. The input characterizing the at least two sound signals includes energy7features and phase features of the at least two sound signals. The phase features include phase difference features between the at least two sound signals. Generating the input characterizing the at least two sound signals includes generating the energy features and the phase features of the at least two sound signals using discrete Fourier transform. Generating the input characterizing the at least two sound signals includes generating the energy features and the phase features of the at least two sound signals using short-time Fourier transform (STFT). Generating the input characterizing the at least two sound signals includes generating the energy features and the phase features of the at least two sound signals using modified discrete cosine transform (MDCT). The neural network is a convolutional neural network, wherein the convolutional neural network includes at least one initial layer that performs a first convolution operation on the energy features and performs a second convolution operation on the phase features, and a subsequent layer that performs a third convolution operation based on outputs from the first convolution operation and the second convolution operation. The at least two sound signals capture multiple sound events emitted from multiple sound sources including the first sound source that emitted the first sound event, the actions include: processing the input characterizing the at least two sound signals using the neural network to generate: (a) the respective score for each of the pluralityof sound event classes indicating predicted classes for the multiple sound events detected in the at least two sound signals, and (b) predicted locations of the multiple sound sources. The predicted location of the first sound source includes a direction of the first sound source relative to one of the at least two audio sensors. The neural network is a convolutional neural network, wherein the convolutional neural network includes one or more initial layers that performs a padded two-dimensional convolution in a channel dimension. The first sound source is a moving sound source, the neural network is a recurrent neural network, the actions include: obtaining the at least two sound signals at one or more additional time points after the first time point; generating the input characterizing the at least two sound signals at the one or more additional time points; processing the input characterizing the at least two sound signals at the one or more additional time points using the recurrent neural network to generate respective predicted locations of the first sound source at the one or more additional time points; and generating tracking information of the first sound source using the respective predicted locations of the first sound source at the first time point and at the one or more additional time points. The predicted location is a two-dimensional location, the at least two sound signals include three sound signals received by three audio sensors at three different locations. The predicted location is a three-dimensional location, the at least two sound signals include four sound signals received by four audio sensors at four different locations.

[0008] Particular embodiments of the subject matter described in this specification can be implemented so as to realize one or more of the following advantages.

[0009] The systems and methods described in this specification are computationally efficient because a single trained neural network is used to perform both sound event classification and sound localization. By using the single trained neural network, the systems and methods can improve both the sound localization and classification capabilities. In some implementations, the systems and methods described in this specification can process the energy features and the phase features of two or more sound signals captured by two or more audio sensors using a neural network to concurrently generate sound event classification and sound localization. In some implementations, the systems and methods described in this specification can perform sound event classification and localization on multiple simultaneous sound events using a neural network that is trained to discriminate between various sounds based on their phase features and energy features. In some implementations, the systems and methods described in this specification can generate tracking information of a moving sound source by predicting locations of the moving sound source at multiple time points, e.g.. by using a recurrent neural network (RNN). In some implementations, the systems and methods can beless computationally expensive than some other approaches. For example, the systems and methods do not need to perform sound separation and sound separation can require significant computation resources.

[0010] The details of one or more embodiments of the subject matter of this specification are set forth in the accompanying drawings and the description below. Other features, aspects, and advantages of the subject matter will become apparent from the description, the drawings, and the claims.BRIEF DESCRIPTION OF THE DRAWINGS

[0011] FIG. 1 is a diagram of an example system.

[0012] FIG. 2 is a diagram of an example signal processing engine.

[0013] FIG. 3 A shows an example of energy features of a sound event.

[0014] FIGs. 3B and 3C show an example of phase features of a sound event.

[0015] FIG. 3D shows an example of phase difference features of sound signals received by two audio sensors.

[0016] FIG. 4 is a flow chart of an example process for jointly performing sound event classification and sound localization.

[0017] FIG. 5 illustrates an example of sound signals of two sound events emitted from two sound sources.|00018| Like reference numbers and designations in the various drawings indicate like elements.DETAILED DESCRIPTION

[0019] FIG. 1 is a diagram of an example system 100.

[0020] A sound event is an event that includes a sound at a single time point or a sound sequence over a period of time in a multi-dimensional space, e.g., a three-dimensional space. A sound event can belong to a sound event class, such as dog barking, human speech, siren, music, etc. A sound source emits a sound event and sound signals of the sound event are captured by the array of two or more audio sensors. A sound signal is a representation of sound, and a sound signal can be a digital sound signal or an analog sound signal. An analog sound signal can use a changing level of electrical voltage to represent sound. A digital sound signal can use a series of binary numbers to represent sound. An analog to digital converter can convert an analog sound signal to a digital sound signal. For example, pulse-code modulation (PCM) is a method used to digitally represent sampled analog sound signals.

[0021] The system 100 includes an array of two or more audio sensors at different locations. Each audio sensor can capture a corresponding sound signal. Although the description below describes two sound signals, similar techniques can apply to a system with three or more audio sensors that captures three or more sound signals.

[0022] For example, the system 100 includes a first microphone 116 and a second microphone 118. A human speaker 112 speaks “Testing 1-2-3”, and thus generates a human speech event 114. The human speaker 112 is a sound source, and the location of the sound source is the location of the human speaker. The human speech event 1 14 arrives at the two microphones 116 and 118 and the first microphone captures a first sound signal 120 while the second microphone captures a second sound signal 122.

[0023] The distance between the sound source and the first audio sensor can be different from the distance between the sound source and the second audio sensor. Thus, there can be a phase shift between the sound signals captured by the first and second audio sensors. For example, the distance dl between the human speaker 112 and the first microphone 116 is shorter than the distance d2 between the human speaker 112 and the second microphone 118. Thus, the second sound signal 122 can arrive at the second microphone 118 later than when the first sound signal 120 arrives at the first microphone 116.

[0024] The system 100 includes a signal processing engine 102. The signal processing engine 102 implements digital signal processing (DSP) techniques to process the first sound signal 120 and the second sound signal 122 captured by the audio sensors 116 and 118. The signal processing engine 102 generates an input 106 characterizing the two sound signals 120 and 122. In some implementations, the system can perform preprocessing on the sound signals, such as resampling or filtering, to focus on specific frequencies or frequency bands of the sound signals, etc.

[0025] FIG. 2 is a diagram of an example signal processing engine 200. The signal processing engine 200 processes two input sound signals 120 and 122 and generates input 106 that includes energy features and phase features of the two sound signals 120 and 122. The signal processing engine 200 can generate energy features and phase features of the two sound signals 120 and 122 using discrete Fourier transform (DFT). The energy features of a sound signal can be the magnitude of an energy spectrum of the sound signal. The phase features of a sound signal can be the phase information included in the DFT of the sound signal.

[0026] For example, the signal processing engine 200 can include a DFT component 202 that generates energy features 204 and phase features 206 of the first sound signal 120. Theenergy features 204 can be the magnitude of the energy spectrum of the first sound signal 120. The DFT component 202 can generate energy features 208 and phase features 210 of the second sound signal 122.

[0027] FIGs. 3 A, 3B, and 3C show an example of energy features and phase features of the human speech event 114. In this example, the signal processing engine 200 computes the energy features and phase features using a short-time Fourier transform (STFT), which evaluates the Fourier transform over a short time window. FIG. 3A shows the magnitude of the energy spectrum of the first sound signal 120. Because the sound signals 120 and 122 originated from the same “human speech” sound event, the energy features of the second sound signal 122 (not shown) is similar to the energy features of the first sound signal 120 in FIG. 3 A. FIG. 3B shows an example of phase features of the first sound signal 120 received by the first audio sensor 116. FIG. 3C shows an example of phase features of the second sound signal 122 received by the second audio sensor 118.

[0028] The phase features in FIG. 3B and FIG. 3C are unwrapped to (-2a, +2a) rad range, and both plots are shown using a logarithmic scale. Because of the highly-varied frequency content of the sound source, the signal fundamental frequencies and their harmonics may not match the DFT frequencies. Thus, there is a DFT leakage between the DFT frequency intervals, and the phase features in FIG. 3B and FIG. 3C may look like static noise or white noise. There can also be a small amount of ambient noise in the sound signals 120 and 122. |00029| Referring back to FIG. 2, because the array of two or more audio sensors are at different locations, there can be a phase shift between the sound signals captured by the audio sensors. In some implementations, the input 106 characterizing the two sound signals 120 and 122 can include energy' features 216 and phase difference features 218 between the two sound signals 120 and 122. The phase difference features 218 can indicate the phase shift between the two sound signals captured by the two audio sensors. The system 100 can perform sound event classification and sound localization based on the energy features 216 and phase difference features 218.

[0030] The signal processing engine 200 can include a combination component 212 that combines the energy features 204 of the first sound signal 120 and the energy features 208 of the second sound signal 122. For example, the energy features 216 included in the input to the neural netw ork 104 can be based on a sum, an average, or other type of combination of the energy features 204 and the energy' features 208. In some implementations, the energy features 216 can be based on the energy features of a single sound signal. For example, the energy feature 216 can be the energy' features 204 of the first sound signal 120 or can be theenergy features 208 of the second sound signal 122. In some implementations, the system 100 can combine the multiple sound signals in the time domain, e.g., taking a sum of the two sound signals 120 and 122, and can generate the energy features included in the input 106 to the neural network 104 by performing DFT on the combined sound signals.

[0031] The signal processing engine 200 can include a subtraction component 214 that calculates the phase difference features 218 by computing a difference between the phase features 206 of the first sound signal 120 and the phase features 210 of the second sound signal 122. In some implementations, the system 100 can include three or more audio sensors, and the signal processing engine can generate phase difference features 218 between each pair of sound signals captured by the three or more audio sensors.

[0032] FIG. 3D shows an example of phase difference features of sound signals received by two audio sensors. FIG. 3D shows the phase difference features between the phase features 206 in FIG. 3B of the first sound signal 120 and the phase features 210 in FIG. 3C of the second sound signal 122. In this example, the difference between d2 and dl in FIG. 1 is about 2.74 cm and the second sound signal 122 was captured about 80 microseconds later than the first sound signal 120. The phase difference features in FIG. 3D show a number of horizontal bands. The system 100 can predict the type of sound event and the location of the sound source based on the horizontal bands in the phase difference features.

[0033] Referring back to FIG. 1, the system 100 includes a neural network 104 trained to jointly perform sound event classification and localization. The system 100 processes the input 106 characterizing the two sound signals using the neural network 104 to generate a predicted class of the sound event and a predicted location of the sound source. For example, the system 100 can process the input 106 characterizing the sound signals 120 and 122 using the neural network 104 to generate a predicted class of "‘human speech” 108 and a predicted location of “11 o’clock direction relative to the second microphone 118”.

[0034] In some implementations, the system 100 can generate a respective score for each of a plurality7of sound event classes. The sound event class with the highest score can be the predicted class for the sound event detected in the two sound signals. For example, the system can have 6 predetermined sound classes, such as dog barking, cat meowing, human speech, siren, music, and car horn. The system can generate a respective score for each sound event class. The sound event class with the highest score can be the “human speech” sound event class. Thus, the predicted class for the sound event detected in the two sound signals 120 and 122 is the “human speech” sound class.

[0035] The number of features in the input 106 can be quite large in order to capture the information of the two or more sound signals 120 and 122. For example, the input 106 can include a large number of phase features to capture the phase difference information with enough granularity and less DFT leakage. The neural network 104 can have a computationally efficient model architecture trained to process the large number of features in the input 106. For example, the neural network 104 can include a convolutional neural network (CNN) model, a dense neural network model, a sequence model such as a gated recurrent unit (GRU), or a combination of these.

[0036] When there are multiple audio sensors, the phase difference features can include phase difference information between each pair of sound signals captured by the two or more audio sensors. For example, when there are n (e.g., n=2, 3, 4, . . . ) audio sensors, the number of pairs of sound signals can be2(n^2y- The input 106 can have 1 +2(-n^2y channels,including one set of energy features, and2^2y sets of phase difference features.

[0037] In some implementations, the neural network 104 can include a CNN model that convolves the energy features and phase features separately in initial layers before combining the actions in later layers. The CNN can include at least one initial layer that performs a first convolution operation on the energy features and performs a second convolution operation on the phase features. The CNN can include a subsequent layer that performs a third convolution operation based on outputs from the first convolution operation and the second operation.

[0038] In some implementations, the inputs to the neural network 104 can be a three- dimensional input and the three dimensions can be time, frequency and channel dimensions. The neural network 104 can include a CNN model that performs convolutions in the time and channel dimensions, instead of performing convolutions in the time and frequency dimensions. For example, the CNN model can include one or more initial layers that performs a padded two-dimensional convolution in the time and channel dimensions.

[0039] The neural network 104 can include a first output layer that generates sound event classification and a second output layer that generates sound localization. The first output layer can be a classification layer that generates a respective score for each of a plurality of sound event classes indicating a predicted class for a sound event detected in the at least two sound signals. For example, the predicted class for the sound event can be the sound class with the highest score.

[0040] The second output layer can be a regression layer that generates a predicted location, e.g., one or more directional vectors, of the sound source that emitted the sound event. Forexample, the regression layer can generate a direction angle (in 2D) or direction angles (in 3D) of the sound source relative to one of the two audio sensors. In some implementations, the regression layer can generate a direction of the sound source relative to another anchor point in the environment, such as an anchor point in between the audio sensors. For example, the range of possible values for the predicted location can be [0, 2rt degrees on a plane defined using the location of an audio sensor or an anchor point, or [0, 2n] degrees by [0, 2TI] degrees in a three-dimensional space defined using the location of an audio sensor or an anchor point.

[0041] In some implementations, the second output layer that generates the sound localization can be a classification layer that generates a distribution over a set of regions in the environment. For example, in a Cartesian system, a two-dimensional plane can include four quadrants, and the second output layer can be a classification layer that generates a distribution of quadrants for the sound source in an environment and the distribution of quadrants can indicate a predicted location of the sound source. For example, the predicted location of the sound signal can be the quadrant with the highest likelihood value.

[0042] In some implementations, the system 100 can be implemented in a computing device, such as a mobile device. For example, the system 100 can use two or more microphones of a computer to capture the sound signals and can process the sound signals on the computer. In some implementations, the system 100 can be implemented on multiple computing devices at the same location or at multiple different locations. For example, the two or more audio sensors, signal processing engine 102, and neural network 104 can be implemented at two or more different devices at the same location. As another example, the system can capture the sound signals using audio sensors at a local computer and can send the sound signals 120 and 122 to a remote computer system, e.g., a cloud server, that implements the signal processing engine 102 and the neural network 104.

[0043] The system 100 can use hardware acceleration or other special -purpose computing devices to implement the operations of the neural network 104. For example, some operations of some layers of the neural network 104 may be performed by highly parallelized hardware, e.g., by a graphics processing unit (GPU) or another kind of specialized computing device. In other words, not all operations of each layer need to be performed by central processing units (CPUs) of the system 100.

[0044] A training system can train the neural network 104 on training data characterizing features of sound signals captured by multiple audio sensors. The training system can beremote from the system 100, e.g., in a data center, or can be at a local computer that deploys the neural network.

[0045] The training data can include training examples. Each training example can include sound features, e.g., energy features and phase features of two or more sound signals captured by two or more audio sensors, and annotation identifying the ground truth sound event class and the ground truth location of the sound source. Once trained, the system 100 can use the neural network 104 to jointly perform sound event classification and localization.

[0046] The system 100 can provide the sound event classification result, e.g., the predicted class 108, and the sound localization result, e.g., the predicted location 110, to a downstream application. In some implementations, the system 100 can use the sound localization result to localize different people who are speaking in a physical space, such as in a meeting room. For example, the system can compare the sound localization result with the location of each person in a meeting room to automatically determine the person who is talking.

[0047] In some implementations, the system 100 can be a hearing aid system and the system can generate hearing aid information based on the sound event classification and sound localization results.

[0048] In some implementations, the system 100 provide the sound event classification and sound localization results to an augmented reality7(AR) system, a smart glass system, a fire alarm system (e.g., identifying a fire alarm event and a location of the fire alarm), a monitoring system, e.g., a baby monitor system that identifies the location of a crying baby7.

[0049] In some implementations, the system 100 can provide the sound event classification and sound localization results to a transcript generation system and the transcript generation system can automatically add the person who is talking to the transcript.

[0050] In some implementations, the system can determine the location of the sound source or can track the location of a moving sound source, and based on the location, the system can point one or more microphone (e.g., a beamforming microphone array) at the sound source such that the system can obtain higher quality sound signals. In some implementations, the system can provide the one or more sound signals and the direction of the sound source to a denoising engine, e.g., a denoising neural network, to reduce the noise in the one or more sound signals along that direction.[00051J FIG. 4 is a flow chart of an example process 400 for jointly performing sound event classification and sound localization. The process will be described as being performed by an appropriately programmed computer system, such as the system 100.

[0052] The system obtains at least two sound signals captured by at least two audio sensors at two different locations at a first time point (402).

[0053] The system generates an input characterizing the at least two sound signals (404). In some implementations, the input characterizing the at least two sound signals can include energy features and phase features of the at least two sound signals. In some implementations, the phase features can include phase difference features between the at least two sound signals.

[0054] In some implementations, the input to the neural network can include phase features generated from each sound signal. For example, instead of calculating the phase difference features 218, the input 106 to the neural network 104 can include phase features 206 of the first sound signal 120 and phase features 210 of the second sound signal 122. For example, the system 100 or the signal processing engine 200 can skip the pre-processing of generating phase difference features, and can use a neural network that is trained to learn the differences in phase features among the various sound signals captured by the audio sensors and to use the differences in phase features to perform sound location and sound classification.

[0055] For a system with a large number of audio sensors that captures a large number of sound signals, using the raw phase features can save the number of sets of phase features included in the input 106. For example, for a system that uses four audio sensors, instead of having 6 sets of phase difference features between each pair of audio sensors, the input 106 can include 4 sets of phase features generated from the respective sound signal.

[0056] In some implementations, the input to the neural network can include the at least two sound signals, e.g., in the time domain rather than in the frequency domain. For example, the input to the neural network can include raw pulse-code modulation (PCM) signals.

[0057] In some implementations, generating the input characterizing the at least tw o sound signals can include generating the energy features and the phase features of the at least two sound signals using discrete Fourier transform (DFT). In some implementations, generating the input characterizing the at least two sound signals can include generating the energy7features and the phase features of the at least two sound signals using short-time Fourier transform (STFT).

[0058] In some implementations, generating the input characterizing the at least tw o sound signals can include generating the energy features and the phase features of the at least tw o sound signals using modified discrete cosine transform (MDCT). For example, the system can generate the energy features and the phase features using overlapping window MDCT. Overlapping window7MDCT is a fully reversible frequency-domain transform which usescosine component coefficients. The energy and phase features for each sound signal can be encoded in a single input channel of the input to the neural network. For a system that has multiple audio sensors, instead of using a set of energy features and a set of phase features, using a single set of MDCT features for each sound signal can save computation and memory.

[0059] The system processes the input characterizing the at least two sound signals using a neural network to generate a prediction at the first time point (406). The prediction includes (a) a respective score for each of a plurality of sound event classes indicating a predicted class for a first sound event detected in the at least two sound signals, and (b) a predicted location of a first sound source that emitted the first sound event. In some implementations, the predicted location of the first sound source can include a direction of the first sound source relative to one of the at least two audio sensors.

[0060] In some implementations, the neural network can be a convolutional neural network, and the convolutional neural network can include at least one initial layer that performs a first convolution operation on the energy features and performs a second convolution operation on the phase features, and a subsequent layer that performs a third convolution operation based on outputs from the first convolution operation and the second convolution operation.

[0061] In some implementations, the neural network can be a convolutional neural network, and the convolutional neural network can include one or more initial layers that performs a padded two-dimensional convolution in a channel dimension.

[0062] In some implementations, the first sound source can be a moving sound source and the neural network can include a recurrent neural network (RNN). The system can obtain the at least two sound signals at one or more additional time points after the first time point. The system can generate the input characterizing the at least two sound signals at the one or more additional time points. The system can process the input characterizing the at least two sound signals at the one or more additional time points using the recurrent neural network to generate respective predicted locations of the first sound source at the one or more additional time points. The system can generate tracking information of the first sound source using the respective predicted locations of the first sound source at the first time point and at the one or more additional time points.

[0063] For example, the neural network 104 can have a sequence model architecture, such as a GRU or a RNN. Sequence models are machine learning models that input or output sequences of data. The sequence model can be configured to process an input generated from a moving sound source and can generate locations of the sound source at multiple time pointsover time. Thus, the system can track the location of the sound source at multiple time points over time.

[0064] In some implementations, the neural network 104 can have a non-sequence neural network architecture. The input to the neural network at the current time point can include some of the intermediate layer activations from a previous inference evaluation of sound signals at the previous time point using the non-sequence neural network. In some implementations, using the non-sequence neural network can operate effectively similar to an RNN.

[0065] In some implementations, the at least two sound signals can capture multiple sound events emitted from multiple sound sources including the first sound source that emitted the first sound event. The system can process the input characterizing the at least two sound signals using the neural network to generate: (a) the respective score for each of the plurality of sound event classes indicating predicted classes for the multiple sound events detected in the at least two sound signals, and (b) predicted locations of the multiple sound sources. For example, the predicted classes for two sound events can be the two sound classes with top two scores that are larger than a threshold value.

[0066] FIG. 5 illustrates an example of sound signals of two sound events emitted from two sound sources. A human speech event emits from the first sound source, e.g., the person 506. A cat meowing event emits from the second sound source, e.g., the cat 508. Two sound signals 502 and 504 capture the two sound events emitted from the two sound sources.

[0067] The system can generate an input characterizing the two sound signals 502 and 504, and the input can include energy features of both sound events and phase features of both sound events. For example, the energy features generated from one or both the two sound signals 502 and 504 can be a combination, e.g., a sum or a weighted sum, of the energy features generated from the human speech event and the energy features generated from the cat meowing event. The phase features, e.g., the phase difference features, generated from the two sound signals 502 and 504 can be a combination, e.g., a sum or a weighted sum, of the phase features, e.g., phase difference features, generated from the human speech event and the phase features, e.g., phase difference features, generated from the cat meowing event.

[0068] The system can process the input characterizing the human speech event and the cat meowing event using a neural network. The system can generate, using the neural network, a respective score for each of the plurality of sound event classes indicating predicted classes for the multiple sound events detected in the at least two sound signals. For example, the system can have 6 predetermined sound classes, such as dog barking, cat meowing, humanspeech, siren, music, and car hom. The system can generate a respective score for each sound event class, such as 0.05 for dog barking, 0.4 for cat meowing, 0.5 for human speech, 0.01 for siren, 0.01 for music, and 0.03 for car hom.

[0069] The system can determine the predicted classes for the multiple sound events detected in the at least two sound signals based on the scores. For example, the system can compare the scores with a threshold value. The sound event class with a score that satisfies the threshold value (e.g.. larger than a threshold value) can be a detected sound event. For example, the threshold value can be 0.35, and the detected sound events can include a cat meowing event and a human speech event that have a corresponding score above 0.35.

[0070] The system can generate a predicted location for each of the plurality of sound event classes. For example, the neural network 104 can include a regression layer for each predetermined sound event class. Each regression layer can generate a predicted direction for the corresponding sound event class. For example, the system can generate a respective predicted direction for each of the 6 predetermined sound classes. The predicted direction for the human speech sound event can be “11 o’clock direction relative to the second microphone 118”. The predicted direction for the cat meowing sound event can be “5 o’clock direction relative to the second microphone 118”.

[0071] In some implementations, the system can use two sound signals received by two audio sensors at two different locations, e.g., along a line, to determine predicted location(s) for one or more sound sources on a plane. In some implementations, the predicted location can be a two-dimensional location, and the at least two sound signals can include three sound signals received by three audio sensors at three different locations. For example, the system can determine a 360-degree localization using three or more audio sensors on a plane. In some implementations, the predicted location can be a three-dimensional location, and the at least two sound signals can include four sound signals received by four audio sensors at four different locations. For example, the system can determine a 360 by 360 degree location in a volume using four or more audio sensors located in a three-dimensional space.

[0072] This specification uses the term “configured” in connection with systems and computer program components. For a system of one or more computers to be configured to perform particular operations or actions means that the system has installed on it software, firmware, hardware, or a combination of them that in operation cause the system to perform the operations or actions. For one or more computer programs to be configured to perform particular operations or actions means that the one or more programs include instructionsthat, when executed by data processing apparatus, cause the apparatus to perform the operations or actions.

[0073] Embodiments of the subject matter and the functional operations described in this specification can be implemented in digital electronic circuitry, in tangibly-embodied computer software or firmware, in computer hardware, including the structures disclosed in this specification and their structural equivalents, or in combinations of one or more of them. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible non-transitory storage medium for execution by, or to control the operation of, data processing apparatus. The computer storage medium can be a machine- readable storage device, a machine-readable storage substrate, a random or serial access memory device, or a combination of one or more of them. Alternatively or in addition, the program instructions can be encoded on an artificially-generated propagated signal, e.g., a machine-generated electrical, optical, or electromagnetic signal, that is generated to encode information for transmission to suitable receiver apparatus for execution by a data processing apparatus.

[0074] The term “data processing apparatus” refers to data processing hardware and encompasses all kinds of apparatus, devices, and machines for processing data, including by way of example a programmable processor, a computer, or multiple processors or computers. The apparatus can also be. or further include, off-the-shelf or custom-made parallel processing subsystems, e.g., a GPU or another kind of special-purpose processing subsystem. The apparatus can also be, or further include, special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit). The apparatus can optionally include, in addition to hardware, code that creates an execution environment for computer programs, e.g., code that constitutes processor firmware, a protocol stack, a database management system, an operating system, or a combination of one or more of them.

[0075] A computer program which may also be referred to or described as a program, software, a software application, an app. a module, a software module, a script, or code) can be written in any form of programming language, including compiled or interpreted languages, or declarative or procedural languages, and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment. A program may, but need not, correspond to a file in a file system. A program can be stored in a portion of a file that holds other programsor data, e.g., one or more scripts stored in a markup language document, in a single file dedicated to the program in question, or in multiple coordinated files, e.g., files that store one or more modules, sub-programs, or portions of code. A computer program can be deployed to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a data communication network.

[0076] As used in this specification, an "‘engine,” or “software engine,” refers to a software implemented input / output system that provides an output that is different from the input. An engine can be an encoded block of functionality, such as a library, a platform, a software development kit (“SDK”), or an object. Each engine can be implemented on any appropriate type of computing device, e.g., servers, mobile phones, tablet computers, notebook computers, music players, e-book readers, laptop or desktop computers, PDAs, smart phones, or other stationary or portable devices, that includes one or more processors and computer readable media. Additionally, two or more of the engines may be implemented on the same computing device, or on different computing devices.

[0077] The processes and logic flows described in this specification can be performed by one or more programmable computers executing one or more computer programs to perform functions by operating on input data and generating output. The processes and logic flows can also be performed by special purpose logic circuitry, e.g., an FPGA or an ASIC, or by a combination of special purpose logic circuitry and one or more programmed computers. |00078| Computers suitable for the execution of a computer program can be based on general or special purpose microprocessors or both, or any other kind of central processing unit. Generally, a central processing unit will receive instructions and data from a read-only memory' or a random access memory' or both. The essential elements of a computer are a central processing unit for performing or executing instructions and one or more memory’ devices for storing instructions and data. The central processing unit and the memory can be supplemented by, or incorporated in, special purpose logic circuitry'. Generally, a computer will also include, or be operatively coupled to receive data from or transfer data to, or both, one or more mass storage devices for storing data, e.g., magnetic, magneto-optical disks, or optical disks. However, a computer need not have such devices. Moreover, a computer can be embedded in another device, e.g., a mobile telephone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a Global Positioning System (GPS) receiver, or a portable storage device, e.g., a universal serial bus (USB) flash drive, to name just a few.

[0079] Computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media and memory’ devices, including by way ofexample semiconductor memory devices, e.g., EPROM, EEPROM, and flash memory devices; magnetic disks, e.g., internal hard disks or removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks.

[0080] To provide for interaction with a user, embodiments of the subject matter described in this specification can be implemented on a computer having a display device, e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor, for displaying information to the user and a keyboard and pointing device, e.g., a mouse, trackball, or a presence sensitive display or other surface by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback, e.g., visual feedback, auditory feedback, or tactile feedback; and input from the user can be received in any form, including acoustic, speech, or tactile input. In addition, a computer can interact with a user by sending documents to and receiving documents from a device that is used by the user; for example, by sending web pages to a web browser on a user’s device in response to requests received from the web browser. Also, a computer can interact with a user by sending text messages or other forms of message to a personal device, e.g., a smartphone, running a messaging application, and receiving responsive messages from the user in return.

[0081] While this specification contains many specific implementation details, these should not be construed as limitations on the scope of any invention or on the scope of what may be claimed, but rather as descriptions of features that may be specific to particular embodiments of particular inventions. Certain features that are described in this specification in the context of separate embodiments can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination. Moreover, although features may be described above as acting in certain combinations and even initially be claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.

[0082] Similarly, while operations are depicted in the drawings in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. In certain circumstances, multitasking and parallel processing may be advantageous. Moreover, the separation of various system modules and components in the embodiments described above should not be understood as requiring such separation in allembodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0083] Particular embodiments of the subject matter have been described. Other embodiments are within the scope of the following claims. For example, the actions recited in the claims can be performed in a different order and still achieve desirable results. As one example, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some cases, multitasking and parallel processing may be advantageous.

[0084] What is claimed is:

Claims

CLAIMS1. A method performed by one or more computers, the method comprising: obtaining at least two sound signals captured by at least two audio sensors at two different locations at a first time point; generating an input characterizing the at least two sound signals; and processing the input characterizing the at least two sound signals using a neural network to generate a prediction at the first time point, wherein the prediction comprises: (a) a respective score for each of a plurality of sound event classes indicating a predicted class for a first sound event detected in the at least two sound signals, and (b) a predicted location of a first sound source that emitted the first sound event.

2. The method of claim 1, wherein the input characterizing the at least two sound signals comprises energy features and phase features of the at least two sound signals.

3. The method of claim 2, wherein the phase features comprise phase difference features between the at least two sound signals.

4. The method of claim 2, wherein generating the input characterizing the at least two sound signals comprises generating the energy features and the phase features of the at least two sound signals using discrete Fourier transform.

5. The method of claim 2, wherein generating the input characterizing the at least two sound signals comprises generating the energy features and the phase features of the at least two sound signals using short-time Fourier transform (STFT).

6. The method of claim 2, wherein generating the input characterizing the at least two sound signals comprises generating the energy features and the phase features of the at least two sound signals using modified discrete cosine transform (MDCT).

7. The method of claim 2, wherein the neural network is a convolutional neural network, wherein the convolutional neural network comprises at least one initial layer that performs a first convolution operation on the energy features and performs a second convolution operation on the phase features, and a subsequent layer that performs a third convolutionoperation based on outputs from the first convolution operation and the second convolution operation.

8. The method of any one of claims 1-7, wherein the at least two sound signals capture multiple sound events emitted from multiple sound sources comprising the first sound source that emitted the first sound event, the method comprises: processing the input characterizing the at least two sound signals using the neural network to generate: (a) the respective score for each of the plurality of sound event classes indicating predicted classes for the multiple sound events detected in the at least two sound signals, and (b) predicted locations of the multiple sound sources.

9. The method of any one of claims 1-7, wherein the predicted location of the first sound source comprises a direction of the first sound source relative to one of the at least two audio sensors.

10. The method of any one of claims 1-7, wherein the neural network is a convolutional neural network, wherein the convolutional neural network comprises one or more initial layers that performs a padded two-dimensional convolution in a channel dimension.

11. The method of any one of claims 1-7, wherein the first sound source is a moving sound source, the neural network is a recurrent neural network, the method comprises: obtaining the at least two sound signals at one or more additional time points after the first time point; generating the input characterizing the at least two sound signals at the one or more additional time points; processing the input characterizing the at least two sound signals at the one or more additional time points using the recurrent neural network to generate respective predicted locations of the first sound source at the one or more additional time points; and generating tracking information of the first sound source using the respective predicted locations of the first sound source at the first time point and at the one or more additional time points.

12. The method of any one of claims 1-7, wherein the predicted location is a two- dimensional location, the at least two sound signals comprise three sound signals received by three audio sensors at three different locations.

13. The method of any one of claims 1-7, wherein the predicted location is a three- dimensional location, the at least two sound signals comprise four sound signals received by four audio sensors at four different locations.

14. A system comprising one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising: obtaining at least two sound signals captured by at least two audio sensors at two different locations at a first time point; generating an input characterizing the at least two sound signals; and processing the input characterizing the at least two sound signals using a neural network to generate a prediction at the first time point, wherein the prediction comprises: (a) a respective score for each of a plurality of sound event classes indicating a predicted class for a first sound event detected in the at least two sound signals, and (b) a predicted location of a first sound source that emitted the first sound event.

15. The system of claim 14, wherein the input characterizing the at least two sound signals comprises energy features and phase features of the at least two sound signals.

16. The system of claim 15, wherein the phase features comprise phase difference features between the at least two sound signals.

17. The system of claim 15, wherein generating the input characterizing the at least two sound signals comprises generating the energy features and the phase features of the at least two sound signals using discrete Fourier transform.

18. The system of claim 15, wherein generating the input characterizing the at least two sound signals comprises generating the energy features and the phase features of the at least two sound signals using short-time Fourier transform (STFT).

19. The system of claim 15, wherein generating the input characterizing the at least two sound signals comprises generating the energy features and the phase features of the at least two sound signals using modified discrete cosine transform (MDCT).

20. One or more non-transitory storage media encoded with instructions that when executed by a computing device cause the computing device to perform operations comprising: obtaining at least two sound signals captured by at least two audio sensors at two different locations at a first time point; generating an input characterizing the at least two sound signals; and processing the input characterizing the at least two sound signals using a neural network to generate a prediction at the first time point, wherein the prediction comprises: (a) a respective score for each of a plurality of sound event classes indicating a predicted class for a first sound event detected in the at least two sound signals, and (b) a predicted location of a first sound source that emitted the first sound event.

Images