WO2024023946A1 - Speech processing device, speech processing method, and speech processing program - Google Patents

Abstract

A first training device (10) calculates a first loss function that becomes smaller as a vector into which a model has quantized the feature amount of speech becomes closer to a context representation which the model has acquired from the feature amount, and a second loss function that becomes smaller as an accuracy with which the model identifies meta information of the speech on the basis of the context representation becomes higher. The first training device (10) updates the parameter of the model such that the first loss function becomes smaller and the second loss function becomes larger.

Description

Audio processing device, audio processing method, and audio processing program

The present invention relates to an audio processing device, an audio processing method, and an audio processing program.

Conventionally, it is known that the accuracy of a subsequent task can be improved by transferring the parameters of a neural network learned through self-supervised learning using audio data to a specific subsequent task (for example, in a non-patent (See Reference 1).

Here, the latter task refers to a task that uses voice as input, such as voice recognition. In self-supervised learning, parameters are learned so that a context representation can be obtained from speech that takes into account previous and subsequent input. Transformer is known as a neural network that can acquire a context representation (see, for example, Non-Patent Document 2).

However, the conventional technology has a problem in that the accuracy of tasks subsequent to self-supervised learning may decrease.

For example, in the technology described in Non-Patent Document 1, the self-supervised learning model for speech may overfit the learning data of the self-supervised learning. In this case, a mismatch occurs between the self-supervised learning model and the data used in the subsequent task, and an effective representation for the subsequent task cannot be obtained.

In order to solve the above-mentioned problems and achieve the purpose, a speech processing device has a first method in which a vector obtained by quantizing speech features by a model becomes smaller as it approaches the context representation acquired by the model from the features. and a second loss function that becomes smaller as the accuracy with which the model identifies meta information of the speech based on the context representation increases, and the first loss function and an updating unit that updates parameters of the model so that the loss function becomes small and the second loss function becomes large.

According to the present invention, it is possible to prevent the accuracy of tasks subsequent to self-supervised learning from decreasing.

FIG. 1 is a diagram showing an example of the configuration of a first learning device. FIG. 2 is a diagram showing an example of the configuration of the second learning device. FIG. 3 is a diagram showing a configuration example of an estimation device. FIG. 4 is a flowchart showing the overall flow of the learning process. FIG. 5 is a flowchart showing the flow of self-supervised learning processing. FIG. 6 is a flowchart showing the flow of relearning processing. FIG. 7 is a flowchart showing the flow of inference processing. FIG. 8 is a diagram showing an example of a computer that executes a learning program.

Embodiments of an audio processing device, an audio processing method, and an audio processing program according to the present application will be described in detail below based on the drawings. Note that the present invention is not limited to the embodiments described below.

[First embodiment]
In the first embodiment, learning (training) of multiple models is performed. The model is, for example, a neural network and includes a speech encoder, a context network, a quantization network, a classification network and an additional network. Details of each network will be described later.

The additional network is a neural network for calculating the final output in the latter task described above. Tasks include classification tasks, generation tasks, prediction tasks, and the like.

In this embodiment, a task that specifically targets audio is targeted. Tasks targeting speech include speech recognition to obtain text from speech, speech classification to classify speech into predetermined types (e.g. speaker attributes, emotions), speaker identification to identify the speaker of speech, etc. It will be done.

Additionally, the process of optimizing model parameters to improve task accuracy is called learning process. Further, the process of actually executing a task using one or more models including additional networks that have been trained through the learning process is called inference process.

The learning process of this embodiment is comprised of two steps: self-supervised learning process and relearning process.

Here, the speech encoder and context network are called a self-supervised learning model. Self-supervised learning models can be used for several different tasks targeting speech. On the other hand, additional networks are models specialized for specific tasks.

In the self-supervised learning process, a self-supervised learning model is trained. In addition, in the relearning process, additional network learning is performed using the self-supervised learning model that has been trained in the self-supervised learning process.

In this embodiment, the first learning device 10 performs self-supervised learning processing. Further, the second learning device 20 performs relearning processing. Further, the inference device 30 performs inference processing. The first learning device 10, the second learning device 20, and the inference device 50 may be realized by different computers, or may be realized by one computer.

The first learning device 10 is an example of a speech processing device. Moreover, any one or more of the first learning device 10, the second learning device 20, and the reasoning device 50 can function as a speech processing device.

The configuration of the first learning device 10 will be explained using FIG. 1. FIG. 1 is a diagram showing an example of the configuration of a first learning device.

As shown in FIG. 1, the first learning device 10 has a set of an acoustic feature sequence X and a classification label l of meta information {(X ₁ , l ₁ ), ..., (X _M , l _M ); l _M ∈{l ¹ ,..., l ^L }} is input as learning data. The classification label l is the correct label.

Here, M is the number of pairs of audio feature series and classification labels included in the learning data, and is an integer of 1 or more. Further, l ^l is the l-th type of classification label. L is the number of types of classification labels prepared, and is an integer of 2 or more.

Further, the meta information is information representing the domain of audio (call center conversation audio, online conference audio, reading audio, etc.), language, gender, etc.

The acoustic features (elements of the acoustic feature series X) are, for example, log Mel filter bank coefficients (FBANK). In addition, acoustic features are not limited to logarithmic mel filter vans, but include MFCC (Mel frequency cepstral coefficient), ΔMFCC (first derivative of MFCC), ΔΔMFCC (second derivative of MFCC), logarithmic power, Δlogarithmic power (logarithmic power first-order differential), etc. Further, the acoustic feature amount may be a sample of raw speech.

Furthermore, the classification label may be represented by an L-dimensional 1-hot vector.

As shown in FIG. 1, the first learning device 10 includes a speech encoder section 11, a context network section 12, a quantization network section 13, a classification network section 14, a classification learning loss calculation section 15, and a context representation learning loss calculation section 16. , a learning parameter updating unit 17, and model information 10a.

The model information 10a is the parameter of the model used by the first learning device 10. Parameters include neural network weights and biases. Furthermore, in the learning process, the model information 10a is updated as appropriate.

_The audio encoder unit 11 calculates a speech intermediate representation vector sequence Z={z ₁ , . . . , z _T } when the audio feature sequence X={x ₁ , . Here, I is the sequence length of the acoustic feature, and is an integer of 1 or more. Further, T is the sequence length of the voice intermediate feature vector sequence, and is an integer of 1 or more.

The audio encoder unit 11 calculates the audio intermediate feature vector sequence Z as shown in equation (1).

Here, SpeechEncoder() (speech encoder) is a function that has the function of a neural network, for example, a convolutional neural network.

θ _se1 is a parameter of the audio encoder and can be learned. θ _se1 is read from the model information 10a.

The context network unit 12 applies masking to the intermediate feature vector sequence Z, which is the output of the audio encoder unit 11, as shown in equation (2).

The context network unit 12 converts the masked intermediate feature vector series (bar above C) into a context expression C={c ₁ , . . . , c _I } as shown in equation (3).

Here, Masking() is a function that performs masking in the time direction.

ContextNetwork() (context network) is a function that has the function of a neural network, and is, for example, the Transformer described in Non-Patent Document 2.

θ _se2 is a parameter of the context network and can be learned. θ _se2 is read from the model information 10a.

The quantization network unit 13 calculates a speech quantization expression vector sequence Q={q ₁ , ..., q _I } from the intermediate feature vector sequence Z that is the output of the audio encoder unit 11, as shown in equation (4). do.

Here, QuantizationNetwork() (quantization network) is a function that has the function of a neural network, and is composed of, for example, a fully connected neural network and a Gumbel softmax function.

The Gumbel softmax function is a differentiable function for propagating the output of a classifier (for example, a fully connected neural network) to a subsequent network. The Gumbel softmax function is described in Reference 1, for example.

θ _qn is a parameter of the quantization network and can be learned. θ _qn is read from the model information 10a.

Reference 1: E. Jang, S. Gu, and B. Poole, “Categorical reparameterization with Gumbel-softmax,” ICLR, 2017.

The classification network unit 14 generates a probability sequence O={o _l1 ,..., o _lL } for the classification label from the context expression vector sequence C that is the output of the context network unit 12, as shown in equations (5) and (6). Calculate. The number of dimensions of the probability sequence O is L. Each element of the probability series O corresponds to each element {l ¹ ,...,l ^L } of the classification label l.

Here, GRL( ) is a function representing a gradient reversal layer (for example, see reference document 1), and is a function that inverts the sign of the gradient during Backward in the error backpropagation method.

ClassNetwork() (classification network) is a function that has the function of a neural network, and is composed of, for example, a fully connected neural network and a softmax function.

θ _qn is a parameter of the quantization network and can be learned. θ _qn is read from the model information 10a.

The classification learning loss calculation unit 15 calculates the classification learning loss L _class for the classification label l as shown in equation (7).

ClassLoss( ) is a function that calculates the loss for identifying the classification label l, for example, cross entropy loss.

The context expression learning loss calculation unit 16 calculates a loss L _context (context expression learning loss) for learning a context expression as shown in equation (8).

ContextLoss() is a function that calculates loss for learning context expressions, for example, Contrastive loss.

Contrastive loss will be explained. Sim() in equation (8) is a function that calculates the similarity between two vectors, and is, for example, a cosine similarity. ^Q (^ above Q) represents the set of negative examples of the quantized vector. τ is a temperature parameter set in advance.

In the numerator in the log of the third side of equation (8), a pair (positive example) of an element of the quantized expression vector sequence Q and an element of the corresponding context expression vector sequence is used. For example, q _t and c _t are a pair of corresponding elements.

On the other hand, in the denominator in the log of the third side of equation (8), a pair (negative example) of an element of the quantized expression vector series Q and an element of the context expression vector series that do not correspond is used. For example, q _t and c _t' (where t≠t') are a pair of non-corresponding elements.

The learning parameter updating unit 17 updates the parameters of the model based on the classification learning loss L _class and the context expression learning loss L _context .

Calculation of the classification learning loss L _class and the loss L _context for learning the context representation is performed for each mini-batch. Therefore, the learning parameter updating unit 17 updates the parameters for each mini-batch.

Assuming that the learning parameters are Θ _se = {θ _se1 , θ _se1 }, Θ _qn = {θ _qn }, Θ _cn = {θ _cn }, the learning parameter updating unit 17 uses equations (9), (10), and ( 11) Update the parameters using the formula.

Here, ε represents the learning rate, β represents the weight for context representation learning loss, and γ represents the weight for classification learning loss. Further, α represents a weight, and the influence of loss is adjusted by changing it significantly each time learning progresses (updating is repeated in mini-batch units).

As shown in equation (5), the function GRL( ) is introduced in the calculation of the classification network unit 14, so the sign of the term with α in equation (11) is inverted.

In Equation (11), the learning parameter is updated so that it becomes smaller as the vector quantized for the input by the model approaches the acquired context representation, and the accuracy of identifying meta information based on the acquired context representation increases. be done.

In this way, the learning parameter updating unit 17 uses the first loss function (context representation learning loss L _context ) and a second loss function (classification learning loss L _class ) that becomes smaller as the accuracy with which the model identifies speech meta information based on the context expression increases. Then, the learning parameter updating unit 17 updates the parameters of the model so that the first loss function becomes smaller and the second loss function becomes larger (Equation (11)). In this case, the learning parameter update section 17 corresponds to a loss function calculation section and an update section.

More specifically, in the parentheses on the right side of equation (11), the parameter by the update unit is transferred from the first term (the term with β) which is the first loss function to the second loss function. This is a loss function (third loss function) obtained by subtracting a second term (term with α) attached with a weight α that increases as the number of times the update of is repeated.

As described above, in this embodiment, in order to prevent overfitting of the model to the learning data in the self-supervised learning process, an adversarial neural network (ANN) is implemented using equation (11). ANN is described in Reference 1.

In this embodiment, the ANN performs learning so that the context network does not identify meta information regarding audio. This allows the context network to obtain a universal representation without overfitting the learning data.

For example, if the classification label l indicates a speech domain, according to this embodiment, the context network will operate robustly even for speech in an unknown domain.

The meta information includes the domain of the voice, the characteristics of the voice (language, etc.), the attributes of the speaker of the voice (gender, age), etc., and is information different from the content of the utterance expressed in text or the like. Moreover, the content of the utterance may be rephrased as the content of information transmitted by voice.

After the parameters θ _se1 , θ _se2 , θ _qn , and θ _cn are updated by the learning parameter update unit 17, the process is further repeated using the updated parameters. Furthermore, when a predetermined condition (for example, the number of repetitions) is satisfied, the iterative process ends.

Here, the context network unit 12 is an example of a context expression calculation unit that inputs voice features into a model and calculates a context expression. Furthermore, the classification network unit 14 is an example of a meta-information label calculation unit that inputs a context expression into a model and calculates a label that specifies audio meta-information. Further, the quantization network unit 13 is an example of a quantization vector calculation unit that inputs a voice feature amount into a model and calculates a quantized vector.

From this, the learning parameter update unit 17 calculates the first loss function so that the vector calculated by the quantization vector calculation unit becomes smaller as it approaches the context expression calculated by the context expression calculation unit, and the meta information It can be said that the second loss function is calculated so that it becomes smaller as the label calculation accuracy of the label calculation unit increases.

The configuration of the second learning device 20 will be explained using FIG. 2. FIG. 2 is a diagram showing an example of the configuration of the second learning device. The second learning device 20 uses the parameters updated by the first learning device 10 to perform learning of a relearning model for performing tasks related to speech. The model in the first learning device 10 is a model that combines a speech encoder, a context network, a quantization network, and a classification network. On the other hand, the relearning model is a model that combines a speech encoder, a context network, and an additional network.

As shown in FIG. 2, a set of an acoustic feature sequence X and a subsequent task label l' is input to the second learning device 20 as learning data. The subsequent task label l' is the correct label.

Here, the subsequent task label l' corresponds to information according to the task, and does not need to indicate meta information.

For example, when the task is speech recognition to obtain text from speech, the subsequent task label l' is the text corresponding to the speech. Further, the subsequent task label l' may indicate meta information like the classification label l. Note that the processing unit of the text corresponding to the voice in the subsequent task label l' may be a phoneme, a character, or a word.

As shown in FIG. 2, the second learning device 20 includes a speech encoder section 21, a context network section 22, an additional network section 23, a subsequent task learning loss calculation section 24, a learning parameter update section 25, and model information 20a.

The model information 20a is parameters of a model trained by the first learning device 10. The model information 20a includes at least parameters θ _se1 and θ _se2 . Furthermore, the model information 20a includes a parameter θ _add of an additional network depending on the task.

Similar to the audio encoder unit 11, the audio encoder unit 21 calculates an intermediate representation vector sequence Z of audio when the audio feature sequence X={x ₁ , . . . , x _I } is given.

The audio encoder unit 21 calculates the audio intermediate feature vector sequence Z as shown in equation (1).

θ _se1 is a parameter of the audio encoder that has been updated by the first learning device 10, and is read from the model information 20a.

Similarly to the context network unit 12, the context network unit 22 converts the intermediate feature vector sequence Z, which is the output of the audio encoder unit 21, into a context expression C as shown in equation (12). However, unlike the context network unit 12, the context network unit 22 does not perform masking.

θ _se2 is a parameter of the context network that has been updated by the first learning device 10, and is read from the model information 10a.

The additional network unit 23 calculates a probability sequence P (sequence of predicted probabilities) for the subsequent task label from the context expression vector sequence C that is the output of the context network unit 22, as shown in equation (13).

The ClassNetwork() (classification network) in equation (13) is different from the classification network of the first learning device 10, and learning is performed in the second learning device 20.

For example, the classification network of the second learning device 20 is a function having the function of a neural network, and is composed of, for example, a bidirectional LSTM and a softmax function.

θ _add is a parameter of the classification network of the subsequent task and can be learned. θ _addn is read from the model information 20a.

The subsequent task learning loss calculation unit 24 calculates the subsequent task learning loss L _down for the subsequent task label l' as shown in equation (14).

Loss( ) is a function that calculates the loss of the subsequent task (for example, classification loss), for example, cross-entropy loss. Note that Loss( ) is changed as appropriate depending on the type of subsequent task (classification task, generation task, prediction task, etc.).

The learning parameter updating unit 25 updates the parameters of the model based on the loss L _down of the subsequent task.

The learning parameter update unit 25 may fix some parameters and update other parameters. For example, the learning parameter updating unit 25 updates the parameter θ _add without updating the parameters θ _se1 and θ _se2 .

The calculation of the loss L _down of the subsequent task is performed for each mini-batch. Therefore, the learning parameter updating unit 25 updates the parameters for each mini-batch.

After the parameters are updated by the learning parameter update unit 25, the process is further repeated using the updated parameters. Furthermore, when a predetermined condition (for example, the number of repetitions) is satisfied, the iterative process ends.

Inference processing using a learned model will be explained using FIG. 3. FIG. 3 is a diagram showing a configuration example of an estimation device. The inference device 50 uses the relearning model to execute a task.

As shown in FIG. 3, the acoustic feature series X is input to the inference device 50 as learning data. For example, the inference device 50 estimates a label corresponding to the acoustic feature sequence X.

As shown in FIG. 3, the inference device 50 includes a speech encoder section 51, a context network section 52, an additional network section 53, and model information 50a.

The model information 50a is the parameters of each model learned by the first learning device 10 and the second learning device 20. The model information 50a includes a learned speech encoder parameter θ _se1 and a learned context network parameter θ _se2 . Furthermore, the model information 50a includes the learned additional network parameter θ _add .

Similar to the audio encoder unit 21, the audio encoder unit 51 calculates an intermediate expression vector sequence Z of audio when the audio feature sequence X={x ₁ , . . . , x _I } is given.

Similarly to the context network unit 22, the context network unit 52 converts the intermediate feature vector sequence Z, which is the output of the audio encoder unit 51, into a context representation C.

The additional network unit 23 calculates a probability sequence P (sequence of predicted probabilities) for the subsequent task label from the context expression vector sequence C that is the output of the context network unit 52.

The additional network unit 53 outputs classification results based on the probability sequence P. The additional network unit 53 may output the probability sequence P, or may output information specifying the subsequent task label corresponding to the element with the largest value among the elements of the probability sequence P.

[Processing of the first embodiment]
The flow of learning processing and inference processing in the first embodiment will be explained using FIGS. 4, 5, 6, and 7.

FIG. 4 is a flowchart showing the overall flow of the learning process. As shown in FIG. 4, first, the first learning device 10 performs preliminary learning of a speech encoder, a context network, a quantization network, and a classification network (step S1).

Next, the second learning device 20 uses the learned speech encoder and context network to learn an additional network (step S2). At this time, it is also possible to relearn the audio encoder and context network.

FIG. 5 is a flowchart showing the flow of self-supervised learning processing. The self-supervised learning process corresponds to the process of step S1 in FIG.

As shown in FIG. 5, first, the first learning device 10 inputs the acoustic feature sequence to the audio encoder and calculates the intermediate expression vector sequence (step S101).

Next, the first learning device 10 applies masking to the intermediate expression vector sequence, inputs it to the context network, and calculates a context expression vector sequence (step S102).

Additionally, the first learning device 10 inputs the intermediate representation vector sequence to the quantization network and calculates a quantized representation vector sequence (step S103).

Next, the first learning device 10 applies GRL to the context expression vector sequence, inputs it to the classification network, and calculates a probability sequence for the classification label of the meta information (step S104).

Then, the first learning device 10 calculates the classification learning loss based on the calculated probability sequence and the correct classification label of the meta information (step S105).

Furthermore, the first learning device 10 calculates a context expression learning loss based on the context expression vector sequence and the quantized expression vector sequence (step S106).

Furthermore, the first learning device 10 updates the parameters of the audio encoder, context network, quantization network, and classification network based on the classification learning loss and context representation learning (step S107).

Here, if the termination condition is satisfied (step S108, Yes), the first learning device 10 terminates the process. On the other hand, if the end condition is not satisfied (step S108, No), the first learning device 10 returns to step S101 and repeats the process using the model whose parameters have been updated.

The termination conditions include, for example, that the process has been repeated a certain number of times, that the amount of parameter updates has converged, etc.

FIG. 6 is a flowchart showing the flow of the relearning process. The relearning process corresponds to the process of step S2 in FIG.

As shown in FIG. 6, the second learning device 20 first inputs the acoustic feature sequence to the audio encoder and calculates the intermediate expression vector sequence (step S201).

Next, the second learning device 20 inputs the intermediate expression vector sequence to the context network and calculates the context expression vector sequence (step S202).

Next, the second learning device 20 inputs the context expression vector sequence to the additional network and calculates a probability sequence for the classification label according to the task (step S203).

Then, the second learning device 20 calculates additional learning loss based on the calculated probability sequence and the correct classification label according to the task (step S204).

Furthermore, the second learning device 20 updates the parameters of the additional network based on the additional learning loss (step S205). At this time, it is also possible to relearn the audio encoder and context network.

Here, if the termination condition is satisfied (step S206, Yes), the second learning device 20 terminates the process. On the other hand, if the end condition is not satisfied (step S206, No), the second learning device 20 returns to step S201 and repeats the process using the model with updated parameters.

The termination conditions include, for example, that the process has been repeated a certain number of times, that the amount of parameter updates has converged, etc.

FIG. 7 is a flowchart showing the flow of inference processing.

As shown in FIG. 7, first, the inference device 50 inputs the acoustic feature sequence to the audio encoder and calculates the intermediate representation vector sequence (step S501).

Next, the inference device 50 inputs the intermediate representation vector sequence to the context network and calculates the context expression vector sequence (step S502).

Subsequently, the inference device 50 inputs the context expression vector sequence to the additional network and calculates a probability sequence for the classification label according to the task (step S503).

Then, the inference device 50 outputs a classification result based on the calculated probability series (step S504).

[Effects of the first embodiment]
As described above, the first learning device 10 has a first loss function that decreases as the vector obtained by quantizing the voice feature amount by the model becomes closer to the context expression acquired by the model from the feature amount; A second loss function is calculated that becomes smaller as the accuracy with which the model identifies audio meta information based on the context expression increases. The first learning device 10 updates the parameters of the model so that the first loss function becomes smaller and the second loss function becomes larger. This prevents the context network from overfitting to the learning data, and prevents the accuracy of tasks subsequent to self-supervised learning from decreasing.

The first learning device 10 converts the first term, which is the first loss function, into a second term, which is weighted with a weight that increases as the number of repeated parameter updates increases. A third loss function is calculated by subtracting , and the model parameters are updated so that the third loss function becomes smaller. As a result, even if the accuracy of the classification network improves as learning progresses, its influence can be reduced.

The first learning device 10 uses the updated parameters to learn a relearning model that executes tasks related to speech. Thereby, subsequent tasks by the additional network can be executed with high accuracy.

The first learning device 10 inputs the voice features into the model, calculates a context expression, inputs the context expression into the model, calculates a label that specifies the meta information of the voice, and inputs the voice features into the model. and calculate the quantized vector. The first learning device 10 calculates a first loss function such that the calculated vector becomes smaller as it approaches the calculated context expression, and calculates a second loss function such that the calculated vector becomes smaller as the calculation accuracy of the label increases. calculate. Thereby, the first learning device 10 can consistently perform calculations using the model and update parameters.

The speech processing device according to the first embodiment provides a specific improvement over the conventional machine learning method as described in Non-Patent Document 1, and is applicable to the technical field related to speech tasks using machine learning models. This shows an improvement in

[System configuration, etc.]
Further, each component of each device shown in the drawings is functionally conceptual, and does not necessarily need to be physically configured as shown in the drawings. In other words, the specific form of distributing and integrating each device is not limited to what is shown in the diagram, and all or part of the devices may be functionally or physically distributed or integrated in arbitrary units depending on various loads and usage conditions. Can be integrated and configured. Furthermore, each processing function performed by each device is realized in whole or in part by a CPU (Central Processing Unit) and a program that is analyzed and executed by the CPU, or by hardware using wired logic. It can be realized as Note that the program may be executed not only by the CPU but also by another processor such as a GPU.

Further, among the processes described in this embodiment, all or part of the processes described as being performed automatically can be performed manually, or the processes described as being performed manually can be performed manually. All or part of this can also be performed automatically using known methods. In addition, information including processing procedures, control procedures, specific names, and various data and parameters shown in the above documents and drawings may be changed arbitrarily, unless otherwise specified.

[program]
In one embodiment, the speech processing device (the first learning device 10, the second learning device 20, or the inference device 50) installs a program that executes the above processing as packaged software or online software on a desired computer. It can be implemented by For example, by causing the information processing device to execute the above program, the information processing device can be made to function as an audio processing device. The information processing device referred to here includes a desktop or notebook personal computer. In addition, information processing devices include mobile communication terminals such as smartphones, mobile phones, and PHSs (Personal Handyphone Systems), as well as slate terminals such as PDAs (Personal Digital Assistants).

Furthermore, the audio processing device can also be implemented as a learning server device that uses a terminal device used by a user as a client and provides services related to the above-mentioned learning processing to the client. For example, a learning server device is implemented as a server device that provides a learning service that takes learning data as input and outputs parameters of a trained model. In this case, the learning server device may be implemented as a Web server, or may be implemented as a cloud that provides services related to the above-mentioned learning processing by outsourcing.

FIG. 8 is a diagram showing an example of a computer that executes a learning program. Computer 1000 includes, for example, a memory 1010 and a CPU 1020. The computer 1000 also includes a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These parts are connected by a bus 1080.

The memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM (Random Access Memory) 1012. The ROM 1011 stores, for example, a boot program such as BIOS (Basic Input Output System). Hard disk drive interface 1030 is connected to hard disk drive 1090. Disk drive interface 1040 is connected to disk drive 1100. For example, a removable storage medium such as a magnetic disk or an optical disk is inserted into disk drive 1100. Serial port interface 1050 is connected to, for example, mouse 1110 and keyboard 1120. Video adapter 1060 is connected to display 1130, for example.

The hard disk drive 1090 stores, for example, an OS 1091, an application program 1092, a program module 1093, and program data 1094. That is, a program that defines each process of the learning device 5 is implemented as a program module 1093 in which computer-executable code is written. Program module 1093 is stored in hard disk drive 1090, for example. For example, a program module 1093 for executing processing similar to the functional configuration of the learning device 5 is stored in the hard disk drive 1090. Note that the hard disk drive 1090 may be replaced by an SSD (Solid State Drive).

Further, the setting data used in the processing of the embodiment described above is stored as program data 1094 in, for example, the memory 1010 or the hard disk drive 1090. Then, the CPU 1020 reads out the program module 1093 and program data 1094 stored in the memory 1010 and the hard disk drive 1090 to the RAM 1012 as necessary, and executes the processing of the embodiment described above.

Note that the program module 1093 and the program data 1094 are not limited to being stored in the hard disk drive 1090, but may be stored in a removable storage medium, for example, and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and the program data 1094 may be stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.). The program module 1093 and program data 1094 may then be read by the CPU 1020 from another computer via the network interface 1070.

Regarding the above embodiments, the following additional notes are further disclosed.

(Additional note 1)
memory and
at least one processor connected to the memory;
including;
The processor includes:
a first loss function in which a vector obtained by quantizing voice features by a model becomes smaller as it approaches a context representation acquired by the model from the features; A second loss function that becomes smaller as the accuracy of identification is higher, and
A speech processing device that updates parameters of the model so that the first loss function becomes smaller and the second loss function becomes larger.

(Additional note 2)
The audio processing device according to Supplementary Note 1, wherein the processor comprises:
A third term is obtained by subtracting a second term, which is weighted with a weight that increases as the number of times parameter updates are repeated, from the first term, which is the first loss function. Calculate the loss function,
A speech processing device that updates parameters of the model so that the third loss function becomes smaller.

(Additional note 3)
The audio processing device according to Supplementary Note 1, wherein the processor comprises:
A speech processing device that uses updated parameters to train a relearning model to perform speech-related tasks.

(Additional note 4)
The audio processing device according to Supplementary Note 1, wherein the processor comprises:
inputting the voice features into the model and calculating a context representation;
inputting the context representation into the model and calculating a label identifying meta information of the speech;
inputting the feature amount of the voice into the model and calculating a quantized vector;
The first loss function is calculated such that the calculated vector becomes smaller as it approaches the calculated context expression, and the second loss function is calculated such that the calculated vector becomes smaller as the calculation precision of the label increases. Audio Processing equipment.

(Additional note 5)
The audio processing device according to Supplementary Note 1, wherein the processor comprises:
A speech processing device that executes the task using the relearning model.

(Additional note 6)
A non-transitory storage medium storing a program executable by a computer to perform audio processing,
The audio processing includes:
a first loss function in which a vector obtained by quantizing voice features by a model becomes smaller as it approaches a context representation acquired by the model from the features; A second loss function that becomes smaller as the accuracy of identification is higher, and
The parameters of the model are updated such that the first loss function becomes smaller and the second loss function becomes larger.

(Supplementary Note 7)
a first loss function in which a vector obtained by quantizing voice features by a model becomes smaller as it approaches a context representation acquired by the model from the features; a second loss function that becomes smaller as the accuracy of identification increases, and updates the parameters of the model so that the first loss function becomes smaller and the second loss function becomes larger. An inference device comprising: an inference unit that performs inference processing regarding speech using a re-learning model that is trained using parameters of the model that has been trained through pre-learning processing.

10 First learning device 10a, 20a, 50a Model information 11, 21, 51 Audio encoder section 12, 22, 52 Context network section 13 Quantization network section 14 Classification network section 15 Classification learning loss calculation section 16 Context representation learning loss calculation Parts 17, 25 Learning parameter updating unit 24 Post-task learning loss calculation unit 23, 53 Additional network unit

Claims (7)

1. a first loss function in which a vector obtained by quantizing voice features by a model becomes smaller as it approaches a context representation acquired by the model from the features; a second loss function that becomes smaller as the accuracy of identification is higher; a loss function calculation unit that calculates a second loss function;
   an updating unit that updates parameters of the model so that the first loss function becomes smaller and the second loss function becomes larger;
   An audio processing device comprising:
2. The loss function calculation unit assigns a weight to the second loss function from a first term that is the first loss function, which increases as the number of times the update unit repeats updating the parameter. calculate the third loss function by subtracting the second term,
   The audio processing device according to claim 1, wherein the updating unit updates parameters of the model so that the third loss function becomes smaller.
3. The audio processing device according to claim 1, further comprising an additional learning unit that uses the parameters updated by the updating unit to learn a relearning model that executes tasks related to audio.
4. a context expression calculation unit that inputs the feature amount of the voice into the model and calculates a context expression;
   a meta information label calculation unit that inputs the context expression into the model and calculates a label specifying meta information of the audio;
   a quantization vector calculation unit that inputs the feature amount of the voice into the model and calculates a quantized vector;
   It further has
   The loss function calculation unit calculates the first loss function such that the vector calculated by the quantization vector calculation unit becomes smaller as it approaches the context expression calculated by the context expression calculation unit, and The audio processing device according to claim 1, wherein the second loss function is calculated so as to become smaller as the label calculation accuracy of the information label calculation unit increases.
5. The audio processing device according to claim 3, further comprising an inference unit that executes the task using the relearning model.
6. An audio processing method performed by an audio processing device, the method comprising:
   a first loss function in which a vector obtained by quantizing voice features by a model becomes smaller as it approaches a context representation acquired by the model from the features; a second loss function that becomes smaller as the accuracy of identification is higher; a loss function calculation step of calculating a second loss function;
   an updating step of updating parameters of the model so that the first loss function becomes smaller and the second loss function becomes larger;
   An audio processing method characterized by having the following.
7. a first loss function in which a vector obtained by quantizing voice features by a model becomes smaller as it approaches a context representation acquired by the model from the features; a second loss function that becomes smaller as the accuracy of identification is higher; a loss function calculation step of calculating a second loss function;
   updating the parameters of the model so that the first loss function becomes smaller and the second loss function becomes larger;
   An audio processing program that causes a computer to execute.

Images