Humans may engage in human-to-computer dialogs with interactive software applications referred to herein as “automated assistants” (also referred to as “digital agents,” “chatbots,” “interactive personal assistants,” “intelligent personal assistants,” “assistant applications,” “conversational agents,” etc.). For example, humans (which when they interact with automated assistants may be referred to as “users”) may provide commands and/or requests to an automated assistant using spoken natural language input (i.e., utterances), which may in some cases be converted into text (or other semantic representation) and then processed. Also, for example, users can additionally or alternatively provide requests by providing textual (e.g., typed) natural language input. An automated assistant responds to a request by providing responsive user interface output, which can include audible and/or visual user interface output.
As mentioned above, many automated assistants are configured to be interacted with via spoken utterances. Spoken utterances are received at a client device via one or more microphones of the client device. For example, each microphone of the client device generates a corresponding audio signal that varies over time in dependence on sound(s) detected at the microphone. The audio signal(s) received via the microphone(s) (and/or frequency domain representations thereof) can be processed (at the client device and/or remote server device(s)) for one or more purposes, such as automatic speech recognition (e.g., converting audio signal(s) to text, phone(s), phoneme(s), and/or other semantic representation).
The client device(s) via which a user interacts with an automated assistant includes an assistant interface that provides, to a user of the client device, an interface for interacting with the automated assistant (e.g., receives spoken and/or typed input from the user, and provides appropriate audible and/or graphical responses), and that interfaces with one or more additional components that implement the automated assistant (e.g., remote server device(s) that process user inputs and generate appropriate responses).
Audio signal(s) that are received via microphone(s) of a client device (and/or frequency domain representations thereof) are often pre-processed prior to being further processed by automatic speech recognition (ASR) component(s) and/or other component(s). Pre-processing of the audio signal(s) (and/or frequency domain representations thereof) can include, for example, dereverberation. Dereverberation of audio data processes the audio data in an attempt to reduce or eliminate reverberation(s) present in the audio data.
Reverberation(s) are reflection(s) of a sound that are created when the sound is reflected off sound-reflective surface(s) of object(s) (e.g., walls, furniture, people), thereby causing reflection(s) of the sound to be created. Reverberation(s) can be present in an audio signal as a corresponding microphone can receive not only the wave-front arriving directly from the source of the sound (e.g., a human speaker) by the shortest path, but can also receive longer path reflection(s) of that wave-front from surrounding object(s). The longer path reflection(s) of a wave-front are time-delayed relative to the wave-front that arrives by the shortest path, and are typically reduced in amplitude relative to the wave-front that arrives by the shortest path. The reflection(s) of a sound in an audio signal generated by a microphone will be dependent on, for example, the size of the room (or other environment), the position and nature of sound-reflective surface(s), the position of the source of the sound, the position of the corresponding microphone, etc.
Although techniques exist for pre-processing audio data to mitigate reverberation (mitigation of reverberation is also referred to herein as “dereverberation”), such techniques can suffer from one or more drawbacks. For example, in pre-processing audio data that includes a spoken utterance, some techniques can require audio data for the entire spoken utterance to be obtained before dereverberation can be performed on the audio data. For instance, some techniques rely on value(s) for dereverberation (e.g., so called “tap” values), where the value(s) cannot be calculated until the audio data for the entire spoken utterance is received. Waiting for audio data for the entire spoken utterance to be obtained can lead to latency in dereverberation, and a resulting latency in use of dereverberated audio data for ASR and/or other purposes—thereby also causing latency in generating a response from an automated assistant (which can rely on the ASR in generating the response). As another example, some techniques can experience a degradation in performance when a source of audio (e.g., a human speaker) is non-stationary during the course a spoken utterance. With such techniques, this and/or other factor(s) can lead to a loss of meaningful audio in dereverberation, resulting in a dereverberated audio signal that cannot be properly processed by ASR component(s) and/or other component(s) to ascertain the true meaning of a spoken utterance. This can lead to poor performance of the ASR and, as a result, poor performance of one or more other automated assistant components that rely on output of the ASR.
Implementations disclosed herein are directed to utilizing an adaptive multichannel technique to mitigate reverberation present in received audio signals, prior to providing corresponding audio data to one or more components for further processing, such as one or more automatic speech recognition (ASR) components of an automated assistant.
Implementations disclosed herein are “adaptive”, in that they utilize a filter, in the reverberation mitigation, that is online and causal. This enables effective dereverberation to be performed on audio data that captures a spoken utterance, and to be performed on portions of the audio data prior to audio data for the entire spoken utterance being received. For example, one or more iterations of updating the filter can occur prior to the entire spoken utterance being spoken by a human speaker, and updated iteration(s) of the filter utilized in dereverberation of audio data that corresponds to portions of the spoken utterance—prior to the entire spoken utterance being spoken by the human speaker. Such online and causal filter techniques described herein can reduce latency in dereverberation (e.g., relative to non-causal techniques), and resultantly reduce latency in use of dereverberated audio data for ASR and/or other purposes—thereby also reducing latency in generating a response from an automated assistant (which can rely on the ASR in generating the response). Such online and causal filter techniques can additionally or alternatively enable adaptation of the filter to changes in a Room Impulse Response (RIR) during the course of a spoken utterance. Changes in RIR can occur during the course of a spoken utterance due to movement of a source of the spoken utterance (e.g., a human speaker) and/or movement of other environmental object(s).
Implementations disclosed herein are “multichannel”, in that a corresponding audio signal is received from each of multiple audio transducers (also referred to herein as “microphones”) of a client device, and the multiple audio signals (e.g., frequency domain and/or time domain representations thereof) are utilized in updating of the filter—and dereverberation occurs for audio data corresponding to each of the audio signals (e.g., frequency domain and/or time domain representations thereof) prior to the audio data being provided to ASR component(s) and/or other component(s). Utilization of multiple channels in updating of the filter can, for example, mitigate negative impact of spectral null(s) that may be present at a given time in audio signal(s) from a subset of (e.g. one of) the microphones, but not all microphones. In other words, while a spectral null may be present at one frequency in the audio signal from a first microphone at a given time, that may not be the case in the audio signal from additional microphone(s) at or near the given time. Accordingly, through utilization of multiple channels, the filter can be updated to reflect the frequency based on the audio signal(s) from the additional microphone(s), despite the spectral null for the frequency in the audio signal from the first microphone.
Audio data that has been dereverberated according to techniques described herein can be provided for ASR and/or other processing. For example, ASR can be performed on the dereverberated audio data by processing the dereverberated audio data over one or more trained acoustic models to generate one or more semantic representations of the dereverberated audio data. The semantic representation(s) can include, for example, phones, phonemes, and/or text. In implementations where the semantic representation(s) include text, the text can optionally be generated based on phones and/or phonemes generated directly based on processing using a trained acoustic model. A trained acoustic model can be, for example, a neural network model such as neural network model including one or more long short-term memory (LSTM) layers, and trained based on a large corpus of audio data. In some implementations, the trained acoustic model can be trained, in whole or in part, on training audio data that is not dereverberated. In those implementations, use of dereverberated audio data (dereverberated according to techniques described herein) with the trained acoustic model can still result in measurable improvements in accuracy and/or robustness of ASR. In some implementations, the trained acoustic model can be trained, in whole or in part, on training audio data that is dereverberated according to techniques described herein. In some of those implementations, use of dereverberated audio data (dereverberated according to techniques described herein) with the acoustic model trained based on dereverberated audio data can result in further measurable improvements in accuracy and/or robustness of ASR. Accordingly, utilization of techniques described herein can result in more robust and/or more accurate detections of features of streams of audio data in various situations, such as environments with reverberation of an audio source due to room acoustics.
In some implementations, techniques described herein are implemented in combination with an automated assistant. For example, audio data that has been dereverberated according to techniques described herein can be provided to an ASR component and/or other component of an automated assistant, and further processing performed based on the dereverberated audio data. Output of the further processing can be utilized to adapt functionality of the automated assistant. For example, further processing performed by an ASR component of the automated assistant can be utilized to determine text that corresponds to a spoken utterance captured in the audio data, and such text utilized by the automated assistant in generating an appropriate audio and/or visual response to the spoken utterance.
In some implementations, dereverberation of audio data according to techniques described herein is performed by one or more processors of a client device that includes the microphones on which the audio data is based. In some of those implementations, the ASR and/or other components to which the dereverberated audio data is provided can be implemented by one or more of the client device processors and/or by processor(s) of remote server(s). For example, in some of those implementations, the dereverberated audio data is transmitted to one or more remote server(s) that perform ASR based on the dereverberated audio data.
Implementations of reverberation mitigation described herein utilize a weighted prediction error (WPE) technique to mitigate reverberations from audio data, while maximizing the meaningful portion(s) of the audio data (e.g., the shortest path portions of audio from source(s)). More particularly, implementations utilize a WPE technique that is adaptive, and that can be utilized with multiple channels of audio data from a microphone array.
Reverberation is generally modeled as the convolution of a room impulse response (RIR) with an original source signal. Particularly when an audio source is speech, such as when a user is providing voice commands to an automated assistant, the speech is correlated because each sound from the audio source occupies a non-instant amount of time. Further, each word, syllable, and/or phrase has some relation to preceding and/or proceeding sounds. Because speech is correlated, it can be beneficial that an adaptive reverberation filter be adapted in a manner to ensure that the resulting dereverberated audio data will be meaningful when provided to one or more other components, such as a speech-to-text module and/or other ASR module(s). Thus, implementations described herein can effectively ignore correlations that occurred less than a threshold time before a frame that is currently being filtered using an adaptive reverberation filter. For example, implementations described herein can update the adaptive reverberation filter at a given iteration using only audio data that occurred at least a threshold time before a frame that is to be filtered with the updated adaptive reverberation filter. For instance, for a current frame that is to be filtered, the adaptive reverberation filter can be updated based on frames that occurred at least n frames before the current frame, where n is greater than 1 (e.g., n can be 2, 3, 4, or 5). Accordingly, one or more frames that immediately precede a given frame being dereverberated may not be utilized to update the adaptive reverberation filter used in the dereverberation, thereby maintaining correlation (in the dereverberated audio data) between the given frame and the immediately preceding frame(s).
Implementations described herein can additionally or alternatively compensate for the variation in the power of the desired source signal during an utterance. Variation in the power of the source signal can occur as a result of the source of the source signal being non-stationary during the utterance (i.e., the source signal can be stronger when the source is closer to the microphones and weaker at a greater distance) and/or as a result of other object(s) in the environment moving during the utterance. Thus, implementations described herein can seek to maximize the ratio of source signal to reverberation to mitigate the risk of a source signal inadvertently being filtered out of the resulting processed audio stream.
As mentioned above, implementations described herein can additionally or alternatively mitigate negative impact of spectral null(s) that may be present at a given time in audio signal(s) from a subset of (e.g. one) the microphones, but not all microphones. By utilizing multiple channels, as described herein, the deleterious effect of spectral nulls may be mitigated. A spectral null is a portion of an audio stream that does not include information for a particular frequency for the given time and can occur when all information is filtered out after processing. In some instances, spectral nulls occur as a result of destructive interference between channels of a multi-channel device (e.g., echoes from one channel cancel the audio data in a different channel). Further, because each microphone of a microphone array receives a slightly different signal due to its unique distance from the source and any reflective surfaces, while a spectral null may be present at one frequency in the audio signal from a first microphone at a given time, that may not be the case in the audio signal from additional microphone(s) at or near the given time. Accordingly, through utilization of multiple channels, audio data for additional microphone(s) can reflect meaningful audio feature(s) absent from audio data for the first microphone (due to the spectral null). Such meaningful audio feature(s) can be used in ASR and can lead to improved performance of the ASR (relative to ASR based on a single channel with a spectral null). Moreover, an adaptive dereverberation filter can be updated to reflect the frequency based on the audio data from the additional microphone(s), despite the spectral null for the frequency in the audio data from the first microphone. This enables effective mitigation of reverberation(s) of the frequency that are present in subsequent portion(s) of the audio data.
The above is provided as an overview of some implementations disclosed herein. Further description of these and other implementations is provided below.
In some implementations, a method performed by one or more processors is provided that includes: receiving a plurality of audio signal streams, wherein each of the audio signal streams is based on output from a corresponding one of a plurality of microphones of a client device that implements an automated assistant interface of an automated assistant. At each of a plurality of iterations during a spoken utterance of a user that is detected at the plurality of microphones and that influences the audio signal streams, the method includes: converting most recent unprocessed portions of the audio signal streams into corresponding frequency domain representations, updating a multi-microphone adaptive reverberation filter utilizing the corresponding frequency domain representations of the audio signal streams for at least one prior iteration of the plurality of iterations, utilizing the updated multi-microphone adaptive reverberation filter in generating reverberation mitigated versions of the corresponding frequency domain representations for the most recent unprocessed portions of the audio signal streams, and providing the reverberation mitigated versions of the corresponding frequency domain representations for further processing by at least one additional component of the automated assistant.
These and other implementations of technology disclosed herein may optionally include one or more of the following features.
In some implementations, converting the most recent unprocessed portions of the audio signal streams into corresponding frequency domain representations further includes applying a fast Fourier transform to each of the unprocessed portions of the audio signal streams to generate the corresponding frequency domain representations of the most recent unprocessed portions of the audio signal streams. In some of those implementations, the corresponding frequency domain representations can each comprise corresponding values for a plurality of frequency bins.
In some implementations, updating the multi-microphone adaptive reverberation filter utilizing the corresponding frequency domain representations of the audio signal streams for at least one prior iteration of the plurality of iterations includes updating the multi-microphone adaptive reverberation filter utilizing the corresponding frequency domain representations of the audio signal streams for an immediately preceding iteration of the plurality of iterations.
In some implementations, at a given iteration, updating the multi-microphone adaptive reverberation filter utilizing the corresponding frequency domain representations of the audio signal streams for at least one prior iteration of the plurality of iterations further includes updating the multi-microphone adaptive reverberation filter utilizing the corresponding frequency domain representations of the audio signal streams for a prior iteration that is at least N iterations prior to the given iteration, and wherein N is greater than one.
In some implementations, the updated multi-microphone adaptive reverberation filter and the corresponding frequency domain representations for the most recent unprocessed portions of the audio signal streams can be each a corresponding matrix, and utilizing the updated multi-microphone adaptive reverberation filter in generating the reverberation mitigated versions of the corresponding frequency domain representations for the most recent unprocessed portions of the audio signal streams can include: generating a conjugate transpose of the multi-microphone adaptive reverberation filter; and generating the reverberation mitigated version of the corresponding frequency domain representations for the most recent unprocessed portions of the audio signal streams based on subtracting, from the corresponding frequency domain representations, a given matrix that is based on the conjugate transpose of the multi-microphone adaptive reverberation filter.
In some implementations, the further processing includes performing automatic speech recognition and the additional component comprises an automatic speech recognition component. In some of those implementations, the automatic speech recognition component can utilize a trained acoustic model in performing the automatic speech recognition. In some of those instances, in performing the automatic speech recognition, the automatic speech recognition component can utilize the trained acoustic model in processing the reverberation mitigated versions of the corresponding frequency domain representations, for a plurality of the iterations, to generate a semantic representation of the reverberation mitigated versions of the corresponding frequency domain representations, for a plurality of the iterations. The semantic representation may comprise text. In some of those instances, the trained acoustic model can be trained based at least in part on training audio data that is not dereverberated. In other instances, the trained acoustic model can be trained based at least in part on training audio data that is dereverberated.
These and other implementations of technology disclosed herein may optionally include one or more of the following features.
In some implementations, another method performed by one or more processors is provided that includes: receiving, during the spoken utterance, at least: a first stream of first channel audio data frames and a second stream of second channel audio data frames, the first stream of the first channel audio data frames being based on output from a first microphone of a client device and the second stream of the second channel audio data frames being based on output from a second microphone of the client device. Prior to completion of the spoken utterance, the method includes: updating a multi-microphone adaptive reverberation filter using both a first subset of one or more of the first channel audio data frames of the first stream and a second subset of one or more of the second channel audio data frames of the second stream, identifying a given first frame of the first channel audio data frames, and a given second frame of the second channel audio data frames, the given first frame and the given second frame received subsequent to the first subset and the second subset, using the updated multi-microphone adaptive reverberation filter in mitigating reverberation in the given first frame and the given second frame; and performing automatic speech recognition based on the reverberation mitigated given first frame and given second frame.
These and other implementations of technology disclosed herein may optionally include one or more of the following features.
In some implementations, the first stream of the first channel audio data frames can be frequency domain representations of a corresponding first channel audio signal stream of the first microphone, and wherein the second stream of the second channel audio data frames are frequency domain representations of a corresponding second channel audio signal stream of the second microphone.
In some implementations, the performing of the automatic speech recognition can include utilizing a trained acoustic model in processing the reverberation mitigated given first frame and given second frame, in generating a semantic representation of the spoken utterance. In some of those implementations, the trained acoustic model can be trained based at least in part on training audio data that is not dereverberated.
In addition, some implementations include one or more processors (e.g., central processing unit(s) (CPU(s)), graphics processing unit(s) (GPU(s), and/or tensor processing unit(s) (TPU(s)) of one or more computing devices, where the one or more processors are operable to execute instructions stored in associated memory, and where the instructions are configured to cause performance of any of the aforementioned methods. Some implementations also include one or more non-transitory computer readable storage media storing computer instructions executable by one or more processors to perform any of the aforementioned methods.
It should be appreciated that all combinations of the foregoing concepts and additional concepts described in greater detail herein are contemplated as being part of the subject matter disclosed herein. For example, all combinations of claimed subject matter appearing at the end of this disclosure are contemplated as being part of the subject matter disclosed herein.
Now turning to
The client device 106 can be, for example: a desktop computing device, a laptop computing device, a tablet computing device, a mobile phone computing device, a computing device of a vehicle of the user (e.g., an in-vehicle communications system, an in-vehicle entertainment system, an in-vehicle navigation system), a standalone interactive speaker, a smart appliance such as a smart television, and/or a wearable apparatus of the user that includes a computing device (e.g., a watch of the user having a computing device, glasses of the user having a computing device, a virtual or augmented reality computing device). Additional and/or alternative client devices may be provided.
Automated assistant 110 engages in human-to-computer dialog sessions with a user via user interface input and output devices of the client device 106. One such example of an output device is speaker(s) of the client device 106, and one such example of a user interface input device is microphone array 105 of client device 106. For example, a user can provide user interface input directed to the automated assistant 110 by speaking spoken utterances that are detected by the microphone array 105. The microphone array 105 includes a plurality of microphones, each configured to detect audio in the environment and generate a corresponding audio signal stream as output. The audio signal stream generated by a given microphone of the microphone array 105 will vary over time in dependence on the audio detected at the given microphone. As one non-limiting example, the microphone array 105 can include two microphones positioned a fixed distance apart, and each of the two microphones can independently detect audio in the environment and can each generate a respective audio signal stream based on its detections. Each microphone in the microphone array 105 may include hardware and/or software components to convert detected audio into a respective audio signal stream.
The automated assistant includes an output module 131, an automated speech recognition (ASR) module 132, a natural language processor 133, a dialog state tracker 134, a dialog manager 135, an audio conversion module 125, and a reverberation reduction module 127. In some implementations, one or more of the engines and/or modules of automated assistant 110 may be omitted, combined, and/or implemented in a component that is separate from automated assistant 110. Also, in some implementations automated assistant 110 can include additional and/or alternative engines and/or modules. Further, as mentioned above, each of the engines and/or modules of automated assistant 110 can be implemented in whole or in part at the client device 106, and/or in whole or in part in remote server device(s).
The ASR module 132 processes audio data (that has been dereverberated by reverberation reduction module 127) over one or more trained acoustic models to generate one or more semantic representations of the dereverberated audio data. The semantic representation(s) can include, for example, phones, phonemes, and/or text. In implementations where the semantic representation(s) include text, the ASR module 132 can include a speech-to-text (STT) module that can optionally generate the text based on phones and/or phonemes generated directly based on processing using a trained acoustic model. A trained acoustic model can be, for example, a neural network model including one or more long short-term memory (LSTM) layers, and trained based on a large corpus of audio data.
As described in more detail below, the automated assistant 110 can utilize semantic representation(s) generated by ASR module 132 in generating responsive output from the automated assistant 110. For example, where audio data captures a spoken utterance of a user, the ASR module 132 can generate a text representation of the spoken utterance, and the automated assistant 110 can generate responsive audible and/or graphical output to provide that is responsive to the spoken utterance, and that is determined based on the spoken utterance. The output module 131 causes the responsive output to be rendered (e.g., audibly and/or graphically) at the client device 106. In some implementations, the responsive output can include textual data and the output module 131 can include a text-to-speech (TTS) module that converts textual data (e.g., natural language responses formulated by automated assistant 110) into computer-generated audio data. The output module 131 can cause the computer-generated audio data to be rendered via one or more speakers of the client device 106.
In some implementations, in generating responsive output from the automated assistant 110, the semantic representation(s) generated by ASR module 132 are provided to natural language processor 133, dialog state tracker 134, dialog manager 135, and/or other components of the automated assistant 110.
Natural language processor 133 (alternatively referred to as a “natural language understanding engine”) of automated assistant 110 processes free-form natural language input and generates, based on the natural language input, annotated output for use by one or more other components of the automated assistant 110. For example, the natural language processor 133 can process natural language free-form input that is textual input that is a conversion, by ASR module 132, of audio data provided by a user via client device 106. The generated annotated output may include one or more annotations of the natural language input and optionally one or more (e.g., all) of the terms of the natural language input.
In some implementations, the natural language processor 133 is configured to identify and annotate various types of grammatical information in natural language input. For example, the natural language processor 133 may include a part of speech tagger (not depicted) configured to annotate terms with their grammatical roles. Also, for example, in some implementations the natural language processor 133 may additionally and/or alternatively include a dependency parser (not depicted) configured to determine syntactic relationships between terms in natural language input.
In some implementations, the natural language processor 133 may additionally and/or alternatively include an entity tagger (not depicted) configured to annotate entity references in one or more segments such as references to people (including, for instance, literary characters, celebrities, public figures, etc.), organizations, locations (real and imaginary), and so forth. The entity tagger of the natural language processor 133 may annotate references to an entity at a high level of granularity (e.g., to enable identification of all references to an entity class such as people) and/or a lower level of granularity (e.g., to enable identification of all references to a particular entity such as a particular person). The entity tagger may rely on content of the natural language input to resolve a particular entity and/or may optionally communicate with a knowledge graph or other entity database to resolve a particular entity.
In some implementations, the natural language processor 133 may additionally and/or alternatively include a coreference resolver (not depicted) configured to group, or “cluster,” references to the same entity based on one or more contextual cues. For example, the coreference resolver may be utilized to resolve the term “there” to “Hypothetical Café” in the natural language input “I liked Hypothetical Café last time we ate there.”
In some implementations, one or more components of the natural language processor 133 may rely on annotations from one or more other components of the natural language processor 133. For example, in some implementations the named entity tagger may rely on annotations from the coreference resolver and/or dependency parser in annotating all mentions to a particular entity. Also, for example, in some implementations the coreference resolver may rely on annotations from the dependency parser in clustering references to the same entity. In some implementations, in processing a particular natural language input, one or more components of the natural language processor 133 may use related prior input and/or other related data outside of the particular natural language input to determine one or more annotations.
In some implementations, dialog state tracker 134 may be configured to keep track of a “dialog state” that includes, for instance, a belief state of a one or more users' goals (or “intents”) over the course of a human-to-computer dialog session and/or across multiple dialog sessions. In determining a dialog state, some dialog state trackers may seek to determine, based on user and system utterances in a dialog session, the most likely value(s) for slot(s) that are instantiated in the dialog. Some techniques utilize a fixed ontology that defines a set of slots and the set of values associated with those slots. Some techniques additionally or alternatively may be tailored to individual slots and/or domains. For example, some techniques may require training a model for each slot type in each domain.
Dialog manager 135 may be configured to map a current dialog state, e.g., provided by dialog state tracker 134, to one or more “responsive actions” of a plurality of candidate responsive actions that are then performed by automated assistant 110. Responsive actions may come in a variety of forms, depending on the current dialog state. For example, initial and midstream dialog states that correspond to turns of a dialog session that occur prior to a last turn (e.g., when the ultimate user-desired task is performed) may be mapped to various responsive actions that include automated assistant 110 outputting additional natural language dialog. This responsive dialog may include, for instance, requests that the user provide parameters for some action (i.e., fill slots) that dialog state tracker 134 believes the user intends to perform. In some implementations, responsive actions may include actions such as “request” (e.g., seek parameters for slot filling), “offer” (e.g., suggest an action or course of action for the user), “select,” “inform” (e.g., provide the user with requested information), “no match” (e.g., notify the user that the user's last input is not understood), a command to a peripheral device (e.g., to turn off a light bulb), and so forth.
The audio conversion module 125 receives a plurality of audio signal streams and converts the audio signal streams into corresponding frequency domain representations. The audio signal streams are received from the microphones of the microphone array 105 of the client device 106. For example, audio conversion module 125 may receive audio signal streams that includes a first channel audio signal stream received from a first microphone of the microphone array 105 and a second channel audio signal stream that is received from a second microphone of the microphone array 105. The audio conversion module 125 can continuously process the first channel audio signal stream and continuously convert the first channel audio signal stream into a corresponding first stream of frequency domain audio data frames. Likewise, the audio conversion module 125 can continuously process the second channel audio signal stream and continuously convert the second channel audio signal stream into a corresponding second stream of frequency domain audio data frames.
In some implementations, in converting a portion of an audio signal stream into a frequency domain representation, the audio conversion module 125 applies a fast Fourier transform (FFT) to the portion of the audio signal stream to generate a frequency domain representation of the portion of the audio signal stream. For example, in generating a stream of frequency domain audio data frames for a channel, the audio conversion module 125 can iteratively apply an FFT to a most recently unprocessed portion of an audio signal stream for the channel. For instance, at a given iteration an X millisecond portion of the audio signal stream can be processed using the FFT to generate a frequency domain audio data frame, at a next iteration the next X millisecond portion of the audio signal stream can be processed using the FFT to generate a next frequency domain audio data frame, etc.
As the audio conversion module 125 processes the audio signal streams into frequency domain audio data frames, the audio data frames can be provided to the reverberation reduction module 127, which determines and applies a multichannel adaptive reverberation filter to the audio data frames. The reverberation reduction module 127 uses the multichannel adaptive reverberation filter to generate reverberation mitigated versions of the audio data frames. As described, the audio data frames provided to the reverberation reduction module 127 include multiple channels from the microphone array 105. As an example, assuming microphone array 105 includes i microphones, the audio data frames can be modeled as:
yi[n]=hi[n]*x[n] (1)
In this instance, x[n] is the source signal and yi[n] is the signal received at the ith microphone at time n. Convolving with the finite impulse response (FIR), represented by hi[n], results in the summation of delayed and attenuated versions of all past source signals arriving at the ith microphone at time n after being reflected off of surface(s) before arriving at the microphone. As a working example, suppose an array includes two microphones spaced a distance apart. Each of the microphones may receive the same intended source signal x[n] via a direct line-of-sight path. Further, each microphone may receive waves that are the source signal reflected off of surface(s). However, because the microphones in the array are at different locations, each may receive the waves at different times relative to the source signal. For example, referring to
Continuing the same example from
Audio conversion module 125 receives an audio signal from each of the microphones of array 105 that includes the original source signal convolved with an impulse response, which results in a sum of all delayed and weakened versions of the source signal. Because the impulse response for an audio data frame includes reflected versions of the source signal previously received by the microphone, the portion identified as being received previously by the microphone (as the source signal) may be mitigated from any future frames because, in the future frame, that portion of the signal is no longer the source signal but instead a reverberation or echo.
Returning to the previous example and referring again to
In order to retain correlation information in the processed audio, techniques described herein can optionally ignore correlations that are shorter than a threshold of correlation. For example, referring again to
The audio data frames for each of the microphones (k, in the above equations), may include, for a given frame for each of the microphones, multiple reverberations of the original audio source. For example, additional walls may be present to reflect the source signal and therefore audio data frame(s) at a given time may include a portion of each of a multiple of reverberations of the audio source. The above equations selectively processes only that part of the audio stream where the delay in signal received at each of the microphones 105 exceeds a threshold. Thus, if the frame in which a portion of audio stream is received at each microphone of the array 105 does not exceed a threshold delay time, that portion of the audio signal will not be filtered out of the processed audio to protect correlation of the rest of the stream. Further, to compensate for the variation in the power of the desired source signal, the ratio of source signal to reverberation can be maximized to better ensure that source signal is not filtered out of the resulting processed audio stream.
In some implementations, spectral nulls in processed audio streams may be compensated for due to the multichannel microphone array 105. A “spectral null” is defined as a frequency of an audio stream where, for a frame, no frequency information is available for a particular frequency range (or frequency bin). Spectral nulls are a common occurrence in multipath channels, where several different channel paths due to reflections can combine in a negative manner at certain frequencies. Multiple microphones and multiple data streams better ensure that the resulting processed audio includes all meaningful information from the original signal source. Thus, for a given audio frame where a spectral null has been introduced, one or more frames from an alternative stream may include the frequency information and may be utilized to process subsequent frames. An audio data frame that is a frequency representation is often represented as values for each of a plurality of frequency bins (F). For example, for a given portion of an audio signal, the audio signal can be converted into an audio data frame that includes 512 frequency bins. If a bin for a stream is null for a particular frame, one or more of the other streams (i.e., a frame from a stream of a different microphone) may not be null for that frequency for the same portion of the audio stream and thus may be utilized to determine frequency information for that time frame and/or may be utilized to update the filter so that those frequencies are filtered out of subsequent frames as unwanted reverberation.
While dereverberation could be performed in the time domain, the filtering in this implementation is done in the frequency domain as represented by
Ŷi,l[n]=Yi,l[n]−Σm=0M-1Σk=0N′−1Wi,m,l*[k]Ym,l[n−k−Δ] (2)
where i represents the channel being dereverberated. Yi,l[n] represents the nth frame of the ST-DFT of subband/for the incoming signal from channel i, Ŵi,m,l[k] is the kth filter tap applied to the lth subband of the mth input channel to produce the ith channel of dereverberated output and Ŷi,l[n] represents the nth frame of the ST-DFT of the lth subband for the processed signal from channel i. M is the number of inputs from microphones and N is the filter length.
In order to retain correlation information in the processed audio, techniques described herein can optionally ignore correlations that are shorter than a threshold of correlation. This is accomplished with the delay Δ is equation (x). For example, referring again to
Combining all channels i=0, 1, . . . M from equation (x) using matrix notation
Ŷl[k]=Yl[k]−ŴlHŶl (3)
In Equation 3, l represents the frequency bin, k represents the the short-time discrete Fourier transform frame index, and Yl[k] is a M×1 vector that contains the kth frame of the lth subband of the short-time discrete Fourier transform of the received signal for each of the multiple microphones.
Ŷl[n]≡[Ŷ0,l[n]Ŷ1,l[n] . . . YM-1,l[n]]T (4)
is a M×1 vector.
{tilde over (Y)}l[n]≡[{tilde over (Y)}0,l{tilde over (Y)}1,l . . . {tilde over (Y)}M-1,l]T (5)
is a MN×1 vector
where
{tilde over (Y)}1,l[n]≡[Yi,l[n−Δ] . . . Yi[n−Δ−(N′−1)]] (6)
The tap matrix of size MN×M is defined as
Wl≡[w0Tw1T . . . WM-1T] (7)
where
wi≡[wi,0wi,1 . . . wi,M−1] (8)
and
wi,j≡[wi,j[0]wi,j[1] . . . wi,j[N′−1]]. (9)
Although the above Equations represent a technique that can be utilized to mitigate reverberation in audio data, the technique relies on value(s) for dereverberation, where the value(s) cannot be calculated until the audio data for the entire spoken utterance is received. For example, the technique requires the entire utterance to be obtained before the tap values of Equation 10 can be calculated and, consequently, before dereverberation can be applied. Waiting for audio data for the entire spoken utterance to be obtained can lead to latency in dereverberation, and a resulting latency in use of a dereverberated audio data for ASR and/or other purposes—thereby also causing latency in generating a response from an automated assistant (which can rely on the ASR in generating the response). Further, because a user may be in motion while providing audio, the RIR may change in the middle of an utterance. To allow for an adaptive filter that is adjusted in real-time to allow for continuous processing of an utterance before the utterance has concluded, the technique represented may be adjusted as follows. For each step, the error term is:
and α is a forgetting factor that impacts the speed of adaptation where 0<α≤1. For example, in some implementations α can be between 0.8 and 1.0. Solving Equation 21 for each step,
In Equation 15, {circumflex over (∧)}l2[k] is the estimate of the received signal power averaged across the M microphones for frame k. This is estimated using a moving average as follows
where r1 and r2 define the limits of the moving average.
Recognizing the Recursive Relationship
R{tilde over (y)}{tilde over (y)},l[k]≡αR{tilde over (y)}{tilde over (y)},l[k−1]+{tilde over (Y)}l[k]{tilde over (Y)}l[k]H (17)
enables the avoidance of taking the inverse of R{tilde over (y)}{tilde over (y)},l[k] at each time step by applying the matrix inversion lemma such that
Thus, the tap update equation can be obtained as
Ŵl[k]=Ŵl[k−1]+Kl[k]ŶlH[k] (21)
The tap values are iteratively updated based on the audio data frames from the audio signals received from a plurality of microphones. Equation 21 can be applied to Equation 10 to generate audio data frames with mitigated reverberation, and the audio data frames. At a given iteration, the filter defined by Equation 21 can be updated based on audio data frames, from multiple microphones, that are earlier in time than the audio data frames being dereverberated, and utilized in Equation 10 to generate reverberation mitigated versions of the audio data frames being dereverberated. This process can be repeated iteratively for each of multiple iterations.
The resulting dereverberated audio data frames have reduced reverberation with the original source signal maintained. The dereverberated audio data may then be provided to one or more components of the automated assistant 110 for further processing. For example, the dereverberated audio data may be provided to the ASR module 132.
At step 405, audio signal streams are received from a plurality of microphones. Each of the audio signal streams is based on output from one of the microphones and may reflect, for example, all or portions of a spoken utterance of a user (direct path and/or indirect (reverberated) portion(s)). For example, the microphones may be a component of a client device, such as client device 106, that implements one or more aspects of an automated assistant. The user may utter one or more phrases to invoke or otherwise interact with the automated assistant and the microphones may sense the audio and provide corresponding audio signal streams to one or more components of the automated assistant, such as audio conversion module 125 (
At step 410, a portion of each of the audio signal streams is converted into a corresponding frequency domain representation. For example, as described herein, audio conversion module 125 may perform a fast Fourier transform on a portion of each of the audio signal streams. For instance, the audio conversion module 125 can convert a most recent unprocessed portion of an audio signal stream from a first microphone into a corresponding frequency domain representation, can convert a most recent unprocessed portion of an audio signal stream from a second microphone into a corresponding frequency domain representation, and so forth for each of the microphones. In some implementations, conversion of the portion of the audio stream may be performed at the client device 106. In some implementations, conversion may be performed remotely from the client device 106. In some implementations, the portion of converted audio may be stored, such as in a buffer, for subsequent processing and/or for subsequent utilization in updating a reverberation mitigation filter. For example, referring again to
At step 415, an adaptive filter is applied to the current portion of the converted audio stream. The filter may be applied as described herein, resulting in mitigation of reverberation in the processed portion. For example, one or more previous portions of the audio stream may include a portion of an utterance of the user. The current portion may include an additional portion of the utterance and additionally include a reverberation of the previous portion of the utterance. Once the filter has been applied, the reverberation portion is mitigated and the additional portion of the utterance (i.e., the meaningful audio from the portion) is preserved. In some implementations, the processed portion of the audio is stored, such as in a buffer and/or otherwise in memory, for utilization in subsequent updating of the audio stream (i.e., for reverberation mitigation of subsequent portions that include a reverberation of the portion of the utterance spoken in the current portion). Application of the adaptive filter may be performed utilizing one or more equations previously described herein.
At step 420, the adaptive reverberation filter is updated based on one or more prior received and converted portions of an audio signal stream. For example, the adaptive reverberation filter can be updated based on prior corresponding frequency domain representations for the audio signal streams, where the prior corresponding frequency domain representations temporally precede the most recently generated frequency domain representations to which the filter is to be applied. The filter can be updated by a component that shares one or more characteristics with the reverberation reduction module 127. In some implementations, reverberation module 127 can update the filter utilizing a portion of the converted audio signal stream immediately preceding the current portion of the audio stream. In some implementations, reverberation module 127 may utilize portions of the audio stream from at least a threshold number of frames prior to the current portion to update the filter. In some implementations, the update step may be omitted, such as in iterations wherein the first portion of the audio data is being processed and the filter has not been initially generated. Thus, in some implementations, updating the filter may include generating a filter that may then be updated during subsequent iterations of the method.
At step 425, reverberation mitigated audio data is provided to one or more additional components for further processing. For example, the reverberation mitigated audio data may be provided to an ASR component, such as an ASR component that includes a STT component. Because the filter is continuously updated based on new audio stream data being received, the updating and filtering process is iterative. Thus, previously processed portions of the stream, which have been reverberation mitigated, may be utilized to mitigate reverberation in subsequent audio portions. In addition to providing the processed audio to one or more components of the automated assistant, the processed portion is provided to the reverberation reduction module 127 for utilization in subsequent filter updates. For example, referring again to
User interface input devices 622 may include a keyboard, pointing devices such as a mouse, trackball, touchpad, or graphics tablet, a scanner, a touchscreen incorporated into the display, audio input devices such as voice recognition systems, microphones, and/or other types of input devices. In general, use of the term “input device” is intended to include all possible types of devices and ways to input information into computing device 610 or onto a communication network.
User interface output devices 620 may include a display subsystem, a printer, a fax machine, or non-visual displays such as audio output devices. The display subsystem may include a cathode ray tube (CRT), a flat-panel device such as a liquid crystal display (LCD), a projection device, or some other mechanism for creating a visible image. The display subsystem may also provide non-visual display such as via audio output devices. In general, use of the term “output device” is intended to include all possible types of devices and ways to output information from computing device 610 to the user or to another machine or computing device.
Storage subsystem 624 stores programming and data constructs that provide the functionality of some or all of the modules described herein. For example, the storage subsystem 624 may include the logic to perform selected aspects of the method described herein, as well as to implement various components depicted in
These software modules are generally executed by processor 614 alone or in combination with other processors. Memory 625 used in the storage subsystem 624 can include a number of memories including a main random access memory (RAM) 630 for storage of instructions and data during program execution and a read only memory (ROM) 632 in which fixed instructions are stored. A file storage subsystem 626 can provide persistent storage for program and data files, and may include a hard disk drive, a floppy disk drive along with associated removable media, a CD-ROM drive, an optical drive, or removable media cartridges. The modules implementing the functionality of certain implementations may be stored by file storage subsystem 626 in the storage subsystem 624, or in other machines accessible by the processor(s) 614.
Bus subsystem 612 provides a mechanism for letting the various components and subsystems of computing device 610 communicate with each other as intended. Although bus subsystem 612 is shown schematically as a single bus, alternative implementations of the bus subsystem may use multiple busses.
Computing device 610 can be of varying types including a workstation, server, computing cluster, blade server, server farm, or any other data processing system or computing device. Due to the ever-changing nature of computers and networks, the description of computing device 610 depicted in
In situations in which certain implementations discussed herein may collect or use personal information about users (e.g., user data extracted from other electronic communications, information about a user's social network, a user's location, a user's time, a user's biometric information, and a user's activities and demographic information, relationships between users, etc.), users are provided with one or more opportunities to control whether information is collected, whether the personal information is stored, whether the personal information is used, and how the information is collected about the user, stored and used. That is, the systems and methods discussed herein collect, store and/or use user personal information only upon receiving explicit authorization from the relevant users to do so.
For example, a user is provided with control over whether programs or features collect user information about that particular user or other users relevant to the program or feature. Each user for which personal information is to be collected is presented with one or more options to allow control over the information collection relevant to that user, to provide permission or authorization as to whether the information is collected and as to which portions of the information are to be collected. For example, users can be provided with one or more such control options over a communication network. In addition, certain data may be treated in one or more ways before it is stored or used so that personally identifiable information is removed. As one example, a user's identity may be treated so that no personally identifiable information can be determined. As another example, a user's geographic location may be generalized to a larger region so that the user's particular location cannot be determined.
While several implementations have been described and illustrated herein, a variety of other means and/or structures for performing the function and/or obtaining the results and/or one or more of the advantages described herein may be utilized, and each of such variations and/or modifications is deemed to be within the scope of the implementations described herein. More generally, all parameters, dimensions, materials, and configurations described herein are meant to be exemplary and that the actual parameters, dimensions, materials, and/or configurations will depend upon the specific application or applications for which the teachings is/are used. Those skilled in the art will recognize, or be able to ascertain using no more than routine experimentation, many equivalents to the specific implementations described herein. It is, therefore, to be understood that the foregoing implementations are presented by way of example only and that, within the scope of the appended claims and equivalents thereto, implementations may be practiced otherwise than as specifically described and claimed. Implementations of the present disclosure are directed to each individual feature, system, article, material, kit, and/or method described herein. In addition, any combination of two or more such features, systems, articles, materials, kits, and/or methods, if such features, systems, articles, materials, kits, and/or methods are not mutually inconsistent, is included within the scope of the present disclosure.
Number | Name | Date | Kind |
---|---|---|---|
8467538 | Nakatani et al. | Jun 2013 | B2 |
9390712 | Yu | Jul 2016 | B2 |
9697826 | Sainath | Jul 2017 | B2 |
10170134 | Markovich Golan | Jan 2019 | B2 |
20160118038 | Eaton | Apr 2016 | A1 |
20160240210 | Lou | Aug 2016 | A1 |
20160322055 | Sainath | Nov 2016 | A1 |
20160358619 | Ramprashad | Dec 2016 | A1 |
Entry |
---|
Kuttruff, H.; Room acoustics. CRC Press; dated 2016. |
Zelinski, R., A Microphone Array with Adaptive Post-Filtering for 1988. |
Bees, D., et al.; “Reverberant Speech Enhancement Using Cepstral Processing,” in Acoustics, Speech, and Signal Processing; ICASSP-91., International Conference on IEEE, pp. 977-980; dated 1991. |
Yoshioka, T., et al.; “Making Machines Understand us in Reverberant Rooms: Robustness Against Reverberation for Automatic Speech Recognition,” IEEE Signal Processing Magazine, vol. 29, No. 6, pp. 114-126, dated 2012. |
Kinoshita, K., et al., et al; “The Reverb Challenge: A Common Evaluation Framework for Dereverberation and Recognition of Reverberant Speech,” in Applications of Signal Processing to Audio and Acoustics (WASPAA), IEEE Workshop on IEEE, pp. 1-4; dated 2013. |
Barker, J. et al; “The Third Chimespeech Separation and Recognition Challenge: Dataset, Task and Baselines,” in Automatic Speech Recognition and Understanding (ASRU); IEEE Workshop on IEEE; pp. 504-511; dated 2015. |
Delcroix, M., et al.; “Linear Prediction-Based Dereverberation With Advanced Speech Enhancement and Recognition Technologies for the Reverb Challenge,” in REVERB Workshop, dated 2014. |
Yoshioka, T., et al., “The ntt Chime-3 System: Advances in Speech Enhancement and Recognition for Mobile Multi-Microphone Devices,” in Automatic Speech Recognition and Understanding (ASRU), Workshop on IEEE, pp. 436-443; dated 2015. |
Miyoshi, M. et al.; “Inverse Filtering of Room Acoustics,” IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 36, No. 2, pp. 145-152, dated 1988. |
Yoshioka, T. and T. Nakatani, “Generalization of Multi-Channel Linear Prediction Methods for Blind Mimo Impulse Response Shortening,” IEEE Transactions on Audio, Speech, and Language Processing; vol. 20, No. 10; pp. 2707-2720; dated Dec. 2012. |
Barry, J. R., et al.; Digital Communication; New York: Springer; dated 2012. |
Yoshioka, T. et al.; “Adaptive Dereverberation of Speech Signals with Speaker-Position Change Detection” in Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) IEEE; pp. 3733-3736; dated 2009. |
Diniz, P. S.; Adaptive Filtering: Algorithms and Practical Implementation. Norwell, MA, USA: Kluwer Academic Publishers; dated 2002. |
Glorot, X., et al.; “Understanding the Difficulty of Training Deep Feedforward Neural Networks,” in Proc. International Conference on Artificial Intelligence and Statistics AISTATS10); Society for Artificial Intelligence and Statistics; dated 2010. |
Number | Date | Country | |
---|---|---|---|
20190272840 A1 | Sep 2019 | US |
Number | Date | Country | |
---|---|---|---|
62637351 | Mar 2018 | US |