BACKGROUND
This invention relates to locating acoustic triggers, and more particularly to selective transmission of audio data based on detected instances utterances of a trigger word or phrase.
One approach to providing a speech-based user interface for a speech-enabled system is to use a device that monitors an acoustic environment waiting for a user to speak a command that can be interpreted by the system. The system may determine when the speaker intends to interface via the interact by determining when the speaker utters a particular word or phrase designed to “wake” the system. Such a word or phrase may be referred to as a “wakeword” or a “trigger word.”
Speech recognition used to determine the words spoken and further automated understand the intent of the speaker may be computationally expensive, and may be beyond the computational capacity of devices that may be located in the acoustic environment being monitored. One approach to addressing the limited computational capacity of such devices is to perform some of the computation on a server coupled to the devices over a data network, for instance over the public Internet. In some such approaches, the devices send audio data to the server only after a speaker utters the wakeword, and the server performs much of the computation required to interpret the speaker's input.
DESCRIPTION OF DRAWINGS
FIG. 1 is a block diagram of a spoken language processing system with an audio user interface device that uses a trigger detector.
FIGS. 2A-E are block diagrams of components of the system shown in FIG. 1.
FIG. 2A is a block diagram of an acoustic front end (AFE). FIG. 2B is a block diagram of a feature extractor. FIG. 2C is a block diagram of a feature analyzer. FIG. 2D is a block diagram of a trigger detector. FIG. 2E is a block diagram of detection logic. FIG. 2F is a block diagram of a spoken language processing server.
FIG. 3 is a diagram of a time delay neural network (TDNN) of a feature analyzer.
FIG. 4 is a diagram of a run-time implementation of the TDNN of FIG. 3.
FIGS. 5A and 5B are schematic representations of a fully connected neural network interconnect and a decomposed interconnect, respectively.
FIG. 6 is a state transition diagram of a Hidden Markov Model (HMM).
FIG. 7 is a graph of an HMM score over time.
FIGS. 8A-C are top layers of a TDNN. FIG. 8A is a top layer used for parameter estimation of the TDNN with large vocabulary input. FIG. 8B is a top layer used in multi-task training using large vocabulary and trigger word training input. FIG. 8C is a top layer retaining a top layer associated with trigger word outputs.
DETAILED DESCRIPTION
Referring to FIG. 1, a spoken language processing system 100 includes an audio user interface device 110 and a spoken language processing system 190, which is generally distant from the device 110 and in data communication with the device over a network, for instance over the public Internet. In the embodiment shown in FIG. 1, the user interface device 110 includes one or more microphones 121, which sense an acoustic environment in which the device 110 is placed. For example, the device 110 may be placed in a living room of a residence, and the microphones acquire (i.e., sense) an acoustic signal in the environment and produce corresponding analog or digital signals, where the acoustic signal may include speech and non-speech sounds. Users in the environment may interact with the system 100. One way for a user to indicate to the system that he or she wishes to interact is to speak a trigger (where “trigger” is used to denote something that initiates a process or reaction), where the trigger may be a predetermined word or phrase (which may be referred to as a “wakeword”, or a “trigger word”) or some other acoustically distinct event. This trigger is detected by the device 110, and upon detection of the trigger at a particular time (e.g., a time instance or interval), the device passes audio data (e.g., a digitized audio signal or some processed form of such a signal) to a spoken language processing server 190. The device 110 selects a part of the audio data corresponding to a time including an interval of the acoustic signal from a starting time and an ending time, for example, based on an estimate of the time that the trigger began in the acoustic signal and based on a determination that input speech in the acoustic signal has ended. This server processes and interprets the user's acoustic input to the device 110 (i.e., the user's speech input) and generally provides a response to the device for presentation to the user. The presentation of the response may in the form of audio presented via a speaker 124 in the device.
One technical problem addressed by the user device 110 is limiting the amount audio data sent to the server 190. It should be recognized that sending the audio data in a continuous manner may consume excessive communication resources as well as computation resources at the server required to process that data. By limiting the sending of audio data to time intervals associated with detection of a trigger, fewer communication and computation resources are consumed. It should also be recognized that in practice, the user device may make errors in detection of the trigger, with both missed actual triggers and false alarms of triggers that are not present. The quality of the user experience is improved when missed triggers are avoided. The amount of avoidable communication and server computation is reduced when false alarms are avoided. Another technical problem may stem from the computation resources at the user device 110 being limited. There may be a need to achieve low error rates within the computation resources available at the device. Increasing computation resources at the user device may reduce error rates, but increases in cost, size, or power consumption needed for such an increase in computation resource may not be desirable or feasible.
The system 100 described below is an example of a system that provides an improved trigger detection error rate within a limited computation capacity (e.g., limits on the number of operations per second and/or the size of memory available to continuously process monitored audio) of the audio user interface device 110. Before addressing the technical features that yield this improvement, which is not merely an improvement achieved by virtue of applying computing resources to the problem, a description of the overall system is provided below.
Continuing to refer to the embodiment shown in FIG. 1, there are several stages of processing that ultimately yield a trigger detection, which in turn causes the device 110 to pass audio data to the server 190. The microphones 121 provide analog electrical signals that represent the acoustic signals acquired by the microphones. These electrical signals are time sampled and digitized (e.g., at a sampling rate of 20 kHz and 16 bits per sample) by analog-to-digital converters 122 (which may include associated amplifiers, filters, and the like used to process the analog electrical signals). As introduced above, the device 110 may also provide audio output, which is presented via a speaker 124. The analog electrical signal that drives the speaker is provided by a digital-to-analog converter 123, which receives as input time sampled digitized representations of the acoustic signal to be presented to the user. In general, acoustic coupling in the environment between the speaker 124 and the microphones 121 causes some of the output signal to feed back into the system in the audio input signals.
An acoustic front end (AFE) 130 receives the digitized audio input signals and the digitized audio output signal, and outputs an enhanced digitized audio input signal (i.e., a time sampled waveform). Referring to FIG. 2A, an embodiment of the signal processor 130 includes multiple acoustic echo cancellers 232, one for each microphone, which track the characteristics of the acoustic coupling between the speaker 124 and each microphone 121 and effectively subtract components of the audio signals from the microphones that originate from the audio output signal. The acoustic front end 130 also includes a directional beamformer that targets a user by providing increased sensitivity to signal that originate from the user's direction as compared to other directions. One impact of such beamforming is reduction of the level of interfering signals that originate in other directions (e.g., measured as an increase in signal-to-noise ratio (SNR)).
In alternative embodiments, the acoustic front end 130 may include various features not described above, including one or more of: a microphone calibration section, which may reduce variability between microphones of different units; fixed beamformers, each with a fixed beam pattern from which a best beam is selected for processing; separate acoustic echo cancellers, each associated with a different beamformer; an analysis filterbank for separating the input into separate frequency bands, each of which may be processed, for example, with a band-specific echo canceller and beamformer, prior to resynthesis into a time domain signal; a dereverberation filter; an automatic gain control; and a double-talk detector.
Referring again to FIG. 1, a second stage of processing converts the digitized audio signal to a sequence of feature values, which may be assembled in feature vectors. A feature vector is a numerical vector (e.g., an array of numbers) that corresponds to a time (e.g., a vicinity of a time instant or a time interval) in the acoustic signal and characterizes the acoustic signal at that time. In the system shown in FIG. 1, a feature extractor 140 receives the digitized audio signal and produces one feature vector for each 10 milliseconds of the audio signal. In this embodiment, the element of each feature vector represents the logarithm of the energy in an audio frequency band (“log frequency band energies” LFBE), the frequency bands (e.g., frequency bands spaced uniformly in a Mel frequency scale) together spanning the typical frequency range of speech. Other embodiments may use other representations of the audio signal, for example, using Ceptral coefficients of Linear Prediction Coding (LPC) coefficients rather than LFBEs.
Referring to FIG. 2B, an embodiment of the feature extractor 140 includes a short time spectral analyzer 242, which accepts the audio signal, performs time windowing, Fourier analysis, and summation of energy over the ranges of the frequency bands. The output of the spectral analyzed 242 passes through a normalizer, which computes a decaying average of the computed feature vectors (optionally limiting the average over period where substantially input energy or speech-like input is received) and subtracting the average from the computed feature vectors to yield normalized feature vectors. Note that because the elements of the feature vectors in this embodiment are logarithms of energy, subtraction is effectively division in the linear energy domain. In some embodiments, the normalizer 244 also accumulates the standard deviation (square root of the variance) along with the average, and divides the elements of the feature vector with the goal of providing zero mean and unit variance elements in each dimension of the feature vectors.
Referring again to FIG. 1, the normalized feature vectors are provided to a feature analyzer 150, which generally transforms the feature vectors to a representation that is more directly associated with the linguistic content of the original audio signal. For example, in this embodiment, the output of the feature analyzer 150 is a sequence of observation vectors, where each entry in a vector is associated with a particular part of a linguistic unit, for example, part of an English phoneme. For example, the observation vector may include 3 entries for each phoneme of a trigger word (e.g., 3 outputs for each of 6 phonemes in a trigger word “Alexa”) plus entries (e.g., 2 entries or entries related to the English phonemes) related to non-trigger-word speech. In the embodiment shown in FIG. 1, feature vectors are provided to the feature analyzer 150 at a rate of one feature vector every 10 milliseconds, and an observation vector is provided as output at a rate of one observation vector every 10 milliseconds. In general, an observation vector produced by the feature analyzer 150 may depend on not only a current feature vector, but may also depend on a history of feature vectors, for example, on 31 most recent feature vectors (e.g., with the output being delayed by 10 vectors relative to the current feature vector, the 31 vectors include 10 vectors in the “future” relative to the delayed time, and 20 frames in the “past” relative to the delayed time).
Various forms of feature analyzer 150 may be used. One approach uses probability models with estimated parameters, for instance, Gaussian mixture models (GMMs) to perform the transformation from feature vectors to the representations of linguistic content. Another approach is to use an Artificial Neural Network (ANN) to perform this transformation. Within the general use of ANNs, particular types may be used including Recurrent Neural Networks (RNNs), Deep Neural Networks (DNNs), Time Delay Neural Networks (TDNNs), and so forth. Yet other parametric or non-parametric approaches may be used to implement this feature analysis. In the embodiment described more fully below, a variant of a TDNN is used.
Referring to FIG. 2C, the feature analyzer 150 of FIG. 1 may include an optional normalizer 252, which transforms each feature vector so that over time, the entries of the feature vectors have zero mean and unit variance. A delay concatenator 254 receives successive feature vectors (e.g., each feature vector being a 20-dimensional real-valued vector) and outputs a delayed concatenation of the feature vectors, in some embodiments formed as the 31 most recent feature vectors (e.g., each concatenated feature vector being a 620-dimensional real-valued vector). Each of these concatenated feature vectors are passed through an artificial neural network (ANN) 256 to produce the output of the feature analyzer 150 as the outputs of the neural network 256. The neural network 256 is configured with configuration data 258, which includes parameters generally referred to as “weights” of the neural network. The configuration data may also include information governing the structure of the network, including the arrangement of nodes and links within a data flow graph representation of the neural network. Embodiments described in this document make use of a particular structure of the neural network 256, in conjunction with corresponding approaches to the determination of the configuration data in a procedure referred to as “parameter estimation” or “training,” to provide high accuracy within the computational constraints of the user interface device 110.
Before continuing with a description of components of the system 100 that may use of the output of the feature analyzer 150 (see FIG. 1 and FIGS. 2D-F), a more detailed description related to the neural network 256 is provided below.
Referring to FIG. 3, the neural network 256 has a multiple layer structure, where inputs are presented at an input layer 310. A concatenation of successive feature vectors form the input layer. It should be recognized that any concatenation of successive feature vectors corresponds to a time interval of the acoustic signal being processed, with that time interval being the overlap of the time intervals corresponding to each of the individual feature vectors (e.g., the time intervals of the sampled signal used to derive the frequency band energies). In particular, to provide an output of the neural network associated with the tth input frame, 10 frames in the “future”, t+1, . . . , t+10, as well as 20 frames in the past, t−20, t−19, . . . , t−1, are used. Because the neural network must wait for the “future” feature vectors, the output is essentially delayed by at least the time associated with 10 frames (e.g., 10 times 10 ms.). The output of the input layer 310 is therefore the 620 entries of the 31 feature vectors. The outputs of the first layer 310 are processed to provide inputs to the second layer 320. As is conventional with such neural networks, each input value to the second layer is formed as a linear weighted combination (i.e., a linear transformation) of the output values of the first layer (generally also including a constant output value). In this example, the second layer has 1,544 units. Therefore, in general, there could be as many as 620×1,544, or almost 1×106 real valued weights defining the linear transformation 315 from the first layer 310 to the second layer 320.
The neural network 256 uses a time-delayed structure, forming what may be referred to as a Time-Delay Neural Network (TDNN), in which groups of input values to the second layer 320 are formed by time delays of application of a same set of weights 316 to successive groups of output values from the first layer. In this embodiment, groups of 100 outputs of the first layer, each group corresponding to 5 consecutive feature vectors of 20 elements each (and corresponding to a time interval of the acoustic signal having an overall duration on which the feature vectors depend), are used to determine 193 inputs to the second layer. In FIG. 3, the weights used to map the first layer's outputs to the second layer's inputs are represented as weights 316 which define the linear transformation 315. In general, as many as 100×193=19,300 real values represent the element 316. These elements applied at 8 time offsets (−18, −14, −10, −6, −4, 0, +4, +8), such that the 5 input frames are centered at these offsets.
The second layer 320 can be considered to be formed of four groups of units, each with 2×193=386 units. These groups each depend on a set of input frames centered at offsets from t of (−16, −8, −2, +6), respectively. For example, the group at offset +6 has two sets of inputs, one set determined from 5 frames centered at +8 (i.e., from +6 to +10) and one set determined from 5 frames centered at +4 (i.e., from +2 to +6). Therefore the group of units at the second layer at offset +6 depends on 9 input frames at offsets +2 to +10, which are centered at offset +6, illustrated by a grouping in a dashed line labeled t+6. Similarly, the group of units at the second layer corresponding to offset −2 depends on the input frames at offsets −6 to +2, the group correspond to offset −8 on input frames at −12 to −4, and the group corresponding to offset −16 on input frames −20 to −12. The overall mapping from input feature vectors to the sets of 193 inputs at the third later are essentially time delays of each other.
In the second layer 320, each of the 1,544 units implements a fixed non-linearity, in this embodiment a fixed sigmoidal nonlinearity, thereby performing an element-wise non-linear transformation of its inputs to form its outputs. The inputs to the third layer 330 are formed as a linear transformation 325 of the outputs of the second layer 320.
Each group of 386 outputs of a group of units of the second layer is used to determine 193 inputs to the third layer according to a set of weights 326. Each set of weights may be represented by 386×193=74498 real values. As with the mapping from the outputs of the first layer to the inputs of the second layer, the same weights 326 are used for each group of 386 outputs. The units of the third layer 330 can be considered as two groups of 386 units, corresponding to offsets from t of +2 and −12. As with the second layer, each unit of the third layer 330 implements a sigmoidal non-linearity.
Repeating the structure at the second layer, each group of 386 outputs of the third layer 330 maps to a set of 193 inputs to the fourth layer 340 with a set of weights 336 defining a linear transformation 335, and the 386 outputs of the fourth layer 340 map to the outputs of the neural network using a set of weights 346 defining a linear transformation 345. In output layer 350 has 20 outputs, and therefore the weights 346 of the linear transformation 345 may be represented by 386×20 real values.
In some embodiments, the outputs of the output layer 350 may be further processed so that they sum to 1.0 by dividing each output by a constant.
Referring to FIG. 4, it may be recognized based on the description above that the neural network 256 may be implemented efficiently by taking advantage of the time delay structure of the network. Because at each time t, the values of the input layer 310 are essentially shifted by one frame (20 values), the required sets of 193 inputs to the second layer previously been computed for all but the first (most recent) set of 5 successive frames (i.e., for offsets t+6 to t+10). Therefore, the two sets of inputs to one group of units of the second layer 320 may be obtained by computing the 193 inputs for the first set of 5 consecutive frames, and using the result of that same computation delayed by 4 frames (i.e., computed using t+2 to t+6). Similarly the inputs for one group of units of the third layer 330 may be computed using the first frames (i.e., from t+2 to t+10) and the result of that computation delayed by 8 frames (i.e., based on input frames t−6 to t+2). Further, one group of inputs at the fourth layer 340 may be computed using the first frames (i.e., t−6 to t+10) and a delay of 14 frames of that computation (i.e., based on frames t−20 to t−4). It should be evident that using this TDNN structure for the neural network 256 greatly reduces the number of arithmetic operations that have to be performed for each input frame as compared to a neural network that does not impose such structure on the weights between the layers of the network.
Referring to FIGS. 5A-B, the number of arithmetic operations per frame at runtime for computation of the output of the neural network 256 shown in FIG. 4 is further reduced by imposing a structure on the sets of weights. For example, considering a set of weights 326 that maps 386 outputs from one layer to 193 inputs to the next layer, in general, such a mapping may be represented using 74498 real valued weights. Referring to FIG. 5B, to further reduce the number of arithmetic operations required, the mapping from the 386 values to the 193 values is formed as a sequence of two mappings: 386 values to 55 values, followed by 55 values to 193 values. Note that the total number of real values representing this sequence of two mappings is 386×55+55×193=31,845, yielding about half as many arithmetic operations to map the 386 values to the 193 values. Therefore, the procedure for processing the input values (i.e., 386 values) to the transformation represented in FIG. 5B includes computing intermediate values (i.e., 55 values) as a linear transformation of the input values, and then computing the output values (i.e., the 193 values) as another linear transformation of the intermediate values. For example, with reference to FIG. 4, this procedure is applied to output values at the second layer 320 to determine the result of applying the weights 326 to those values. Recall that because of the time-delay structure shown in FIG. 4, the inputs to the procedure (i.e., the 386 values at layer 320) are formed by application of the weights 316 at the previous layer to input corresponding to frames t+6 to t+10 (and the associated time interval of the acoustic signal) and a time delay (by 4 frames) of application of those weights 316 to input corresponding to frames t+2 to t+6 (corresponding to a prior time interval of the acoustic signal).
Versions of this procedure are performed at each layer 315, 325, 335, and 345 to provide an efficient implementation of the overall transformation illustrated in FIG. 3. It should be recognized that the particular numbers of inputs and outputs and intermediate value, the time delays, the number of layers, and any other parameters specified above are merely examples used in the embodiment that is described in detail, and other values of such parameters may be used in other embodiments.
In the computational structure shown in FIG. 4, each of the groups of weights 316, 326, 336 are represented in a similar decomposition with 55 intermediate values. In the specific example shown in FIG. 4, there are a total of about 87,525 weights, and a similar number of arithmetic operations required for processing each input frame using the neural network.
Referring back to FIG. 1 and FIGS. 2D-F, the 20 outputs of the neural network 256, which are the outputs of the feature analyzer 150 are provided to the trigger detector 160. Various types of trigger detectors may be used. One class of trigger detectors is configured to locate instances of a pattern in the successive sets of outputs of the feature analyzer 150. One specific type of such a sequential pattern detector uses a Hidden Markov Model (HMM), however other types may use alternatives such as dynamic time warping, dynamic programming.
Referring to FIG. 2D, in this embodiment the trigger detector uses a Hidden Markov Model (HMM) module 262, and output of the HMM module 262 is provided to detection logic 266. The HMM module 262 is configured with HMM configuration data 264 and the detection logic 266 is configured with detector configuration data 267. Referring to FIG. 6, the HMM configuration data 264 specifies an HMM 600 that represents background (i.e., non-trigger-word) input as well as trigger word input, with this example illustrating a trigger word “Alexa”. The background uses two states of the HMM, one associated with speech and one with non-speech. The trigger word portion represents a sequence of phonemes of the trigger word, with each phoneme being represented by a sequence of three states. It should be recognized that different structures of HMM may be used, for example, with different number and interconnection of states for each phoneme, different states arrangements for non-trigger-word input, and so forth.
At runtime, the HMM module 262 uses a Viterbi algorithm approach, which provides at each time (i.e., input frame) a score associated with each state, with the score at a state at a time representing a likelihood of the most likely sequence of states that ends at that state at that time. These scores can be scaled such that at each time the scores sum to 1.0, and they can be considered to be probabilities. Referring to FIG. 7, a likelihood ration of a probability 710 for the last state of the trigger word relative to the most likely state of the background evolves (i.e., goes up and down). This state value is one of the outputs for the HMM model 262 to the detection logic 266.
Referring to FIG. 2E, the detection logic 266 uses the evolution of the probability shown in FIG. 7 to determine when a candidate trigger word occurs. Generally, when the probability in FIG. 7 reaches a local maximum above a threshold 715, the detection logic 266 evaluates a detection score. More specifically, the detection logic includes a peak detector and a state machine that is used to select the candidate peaks for further evaluation. The detection logic 266 makes use of various quantities determined based on the processing by the HMM module 262, which are passed to a traceback and secondary feature assembly component of the detection logic. Based on the Viterbi processing of the HMM module, a particular ending time 720 for the trigger word is associated with a corresponding starting time 725 making use of a “traceback” data structure computed during the Viterbi processing. In some examples, the detection logic waits for a time 730 to determine whether there is a higher peak, such as at time 740, and if there is such a higher peak, it is chosen rather than the earlier peak. Furthermore, the HMM module provides sequence of states from the starting time to the ending time. Also, the HMM module provides score of the starting time to ending time remaining in the background states of the HMM. Generally, the detection logic receives these various computed quantities corresponding to the trigger word ending time and performs a scoring or detection decision based on the inputs. In some embodiments, the detection logic is implemented using a Support Vector Machine (SVM), which is configured according to parameters of the detector configuration data 267.
Referring back to FIG. 1, the communication interface receives an indicator part of the input (e.g., the frame number) corresponding to the identified trigger. Based on this identified part of the input, the communication interface 170 selects the part of the audio data (e.g., the sampled waveform) to send to the server 190. In some embodiments, this part that is sent starts at the beginning of the trigger, and continues until no more speech is detected in the input, presumably because the user has stopped speaking. In other embodiments, the part corresponding to the trigger is omitted from the part that is transmitted to the server. However, in general, the time interval corresponding to the audio data that is transmitted to the server depends on the time interval corresponding to the detection of the trigger (e.g., the trigger starts the interval, ends the interval, or is present within the interval).
Referring to FIG. 2F processing at the spoken language server 190 may include various configurations for processing the acoustic data (e.g., the sampled audio waveform) received from the audio interface device 110. For example, a runtime speech recognizer 281 uses an acoustic front end 282 to determine feature vectors from the audio data. These may be the same feature vectors computed at the interface device 110, or may be a different representation of the audio data (e.g., different numbers of features, different number per unit time, etc.). A speech recognition engine 282 processes the feature vectors to determine the words in the audio data. Generally, the speech recognizer 281 attempts to match received feature vectors to language phonemes and words as known in the stored acoustic models 283 and language models 285. The speech recognition engine 284 computes recognition scores for the feature vectors based on acoustic information and language information and provides text as output. The speech recognition engine 284 may use a number of techniques to match feature vectors to phonemes, for example using Hidden Markov Models (HMMs) to determine probabilities that feature vectors may match phonemes. Sounds received may be represented as paths between states of the HMM and multiple paths may represent multiple possible text matches for the same sound.
Following processing by the runtime speech recognizer 281, the text-based results may be sent to other processing components, which may be local to the device performing speech recognition and/or distributed across data networks. For example, speech recognition results in the form of a single textual representation of the speech, an N-best list including multiple hypotheses and respective scores, lattice, etc. may be sent to a natural language understanding (NLU) component 291 may include a named entity recognition (NER) module 292, which is used to identify portions of text that correspond to a named entity that may be recognizable by the system. An intent classifier (IC) module 294 may be used to determine the intent represented in the recognized text. Processing by the NLU component may be configured according to linguistic grammars 293 and/or skill and intent models 295. After natural language interpretation, a command processor 296, which may access a knowledge base 297, acts on the recognized text. For example, the result of the processing causes an appropriate output to be sent back to the user interface device for presentation to the user.
The description above focuses on the runtime operation of the audio interface device 110 shown in FIG. 1. As described above, the device 110 is configured with various data, including the neural network configuration data 258, the HMM configuration data 264, and the detector configuration data 267. The neural network configuration data includes the weights 316, 326, 336, 346 of the neural network, as discussed above. In general, the configuration data for the device 110 is computed on a data processing system that is separate from the device 110 based on representative input that includes examples of the trigger word to be detected. The configuration data may be loaded into the device before deployment or downloaded to the device 110 from time to time from a server.
Turning specifically to the procedure for determining the neural network weights, the input to the procedure include data that includes a sequence of feature vectors, and for each feature vector, an indicator of the HMM state at that time or a distribution over the state of the HMM at that time. Note that the indicator of the state and/or the distribution over the states are obtained using conventional HMM parameter estimation techniques, for example, using a Forward-Backward training procedure.
The determination of the weights of the neural network generally follows a sequence of stages: a) initialization; b) multi-task learning; c) weight decomposition; and d) further multi-task learning.
Referring to FIG. 8A, in the initialization stage (a), weights of the neural network are determined using an output layer 850 with outputs corresponding to state of phonemes for a large vocabulary speech recognizer. Here the layer is illustrated as having 120 units, which may correspond to 40 distinct phonemes with 3 states each, however is should be understood that a larger or smaller number of output units may be used, for example, if more phonemes (e.g., context-dependent phonemes) are used. The linear transformation 845 from the fourth layer 340 to the output layer is defined in terms of a set of weights 846. As introduced above, the input to the large vocabulary training procedure includes paired neural network inputs and outputs. The weights 846 are determined using one of a variety of parameter estimation procedures, such as a Back Propagation gradient-based estimation procedure using a cross-entropy loss criterion.
Referring to FIG. 8B, using the weights 846 of the final linear transformation 845, as well as the weights of the transformations 315, 325, 335 other layers (not shown in FIGS. 8A-C) of the remainder of the neural network, which have been initialized using the large vocabulary training data, a multi-task training stage (b) adds weights 346 to the final transformation 845B, providing in the output layer both output units corresponding to units for large vocabulary recognition as well as units for the trigger detection task. The parameters of the neural network are further updated using training data that includes both data for the large vocabulary task as well as training data for the trigger detection task. The parameter estimation procedure uses a weighted cross-entropy loss criterion, for example, weighting the cross-entropy for the trigger detection task at 0.9 and 0.1 for the large vocabulary task.
Referring to FIG. 8C, after the multi-task training stage, the linear transformations between layers are decomposed as illustrated in FIG. 5B in a decomposition stage (c). Each linear transformation is represented as a matrix transformation, and a singular value decomposition (SVD) is used to approximate the full matrix as a product of lower rank matrices. In particular, if the linear transformation from N outputs for one layer to M inputs to the next layer is represented as an N×M matrix, this matrix is decomposed into a product of an N×R matrix and an R×M matrix using a reduced rank approximation based on a singular value decomposition of the N×M matrix. The values of the two matrices are used as initial weights for the transformation of the form illustrated in FIG. 5B.
Having initialized the decomposed weights for the transformation in the neural network, a further stage (d) of multi-task training is performed using large vocabulary and trigger word training data. Finally, as illustrated in FIG. 8C, only the weights 346 of the final linear transformation 345 are retained for the neural network configuration data 258, which is used in the runtime operation of the audio user interface device 110 as described above.
Note that in the description above, particular numbers of concatenated frames, particular time delays, particular numbers of units in layers of the neural network, and so forth, are provided by way of illustrative example. However, it must be recognized that these are only exampled. Different numbers of layers, units in each layer, size of the SVD decomposition (e.g., R above), and so forth may be chosen to achieve selected tradeoffs between computation resources (processor instructions per second, size of memory) required and achieved accuracy.
Furthermore, although described in the context of an interface device that both has the microphones integrated into the device and a processor for execution of instructions that implement the processing for detecting the triggers, the microphones for acquiring the acoustic signals may be separate from the interface device, for example, providing an analog or digitized audio signal over a communication link (e.g., a radio link or a data communication link). Furthermore, some or all of the function of the spoken language processing server 190 may be integrated into the device 110, while still benefitting from the technical advantage of limiting the amount of the audio data that must be processed by such function. For example, certain computation resources in such a device may be in a low-power mode until the trigger is detected when they are used for the spoken language processing of the audio data.
The approaches described above may be implemented in software, in hardware, or using a combination of software and hardware. For example, the software may include instructions stored on a non-transitory machine readable medium that when executed by a processor, for example in the user interface device, perform some or all of the procedures described above. Hardware may include special purpose circuitry (e.g., Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs) and the like) for performing some of the functions. For example, some of the computations for the neural network transformations may be implemented using such special purpose circuitry.
It is to be understood that the foregoing description is intended to illustrate and not to limit the scope of the invention, which is defined by the scope of the appended claims. Other embodiments are within the scope of the following claims.