Model learning device, model learning method, program

Abstract

The present invention provides a model training device including an utterance feature reconfiguration model training unit that performs masking by randomly selecting a portion of an utterance feature sequence, which is a sequence of utterance features corresponding to each utterance of a subject speaker, and replacing the selected portion with predetermined masking information, trains an utterance feature reconfiguration model, which is a neural network model for estimating the utterance features in the masked utterance feature sequence, and outputs the trained utterance feature reconfiguration model as an unsupervised, pretrained model.

Description

Technical Field

[0001] The present disclosure relates to a model learning device for learning a pre-trained model, a satisfaction estimation device for estimating a dialogue satisfaction degree and an utterance satisfaction degree based on an estimation model learned by the pre-trained model, a model learning method, a satisfaction estimation method, and a program.

Background Art

[0002] In dialogue, there is a need for a technique to estimate the satisfaction degree of the target speaker with respect to the entire dialogue (hereinafter referred to as "dialogue satisfaction degree") and the satisfaction degree of the target speaker for each utterance of the target speaker (hereinafter referred to as "utterance satisfaction degree"). Satisfaction degree is a categorical category indicating whether the speaker expresses satisfaction or dissatisfaction, and for example, refers to three levels such as satisfied, normal, and dissatisfied. The most typical application of this technique is to use the dialogue as a call center call and the target speaker as a customer, that is, to estimate the customer satisfaction degree in a call center call. For example, by aggregating the estimation results of the customer's dialogue satisfaction degree for each operator, operator evaluation can be automated, and by collecting only the utterance intervals where the utterance satisfaction degree is "dissatisfied" and performing speech recognition and text analysis, the problems of the service can also be analyzed. Note that such a technique can be applied not only to call center calls but also to all dialogues conducted face-to-face / non-face-to-face among multiple speakers.

[0003] A technique for estimating dialogue satisfaction degree and utterance satisfaction degree from dialogue (hereinafter referred to as the prior art) is disclosed in Patent Document 1. In the prior art, after extracting a feature vector (hereinafter referred to as an utterance feature amount) including one or more of prosodic features, dialogue features, and linguistic features for each utterance of the target speaker, the estimation results of the utterance satisfaction degree and the dialogue satisfaction degree are obtained by inputting them into a model that simultaneously estimates the utterance satisfaction degree and the dialogue satisfaction degree. The point of the prior art is to obtain an estimation model that utilizes the relationship between the utterance satisfaction degree and the dialogue satisfaction degree and hierarchically performs multi-task learning to improve the estimation accuracy of the utterance satisfaction degree and the dialogue satisfaction degree.

Prior Art Documents

Patent Documents

[0004] [Patent Document 1] Patent No. 6852161 [Overview of the project] [Problems that the invention aims to solve]

[0005] Conventional estimation models use deep neural networks, such as Recurrent Neural Networks (RNNs). Training these estimation models requires a large amount of labeled training data. Training data represents a pair of input features and labels that represent the correct values ​​of the information to be estimated. In conventional techniques, this refers to a sequence of utterance features in a dialogue, a sequence of correct values ​​of utterance satisfaction in that dialogue, and the correct value of dialogue satisfaction.

[0006] However, preparing a large amount of labeled training data is extremely costly. This is because the correct values ​​for utterance satisfaction and dialogue satisfaction must be manually labeled by humans after listening to the dialogues. For practical purposes, estimation models must be trained using only a small amount of labeled training data. However, in that case, there is a risk that the estimation accuracy of utterance satisfaction and dialogue satisfaction will decrease.

[0007] Therefore, this disclosure aims to provide a learning device for unsupervised pre-trained models that can construct highly accurate estimation models using a small amount of labeled training data. [Means for solving the problem]

[0008] The model learning device of this disclosure includes a speech feature reconstruction model learning unit.

[0009] The speech feature reconstruction model learning unit randomly selects a portion of the speech feature sequence, which is a sequence of speech features corresponding to each utterance of the target speaker, and replaces it with predetermined masking information to mask it. It then learns a speech feature reconstruction model, which is a neural network model that estimates the speech features of the masked speech feature sequence, and outputs the learned speech feature reconstruction model as an unsupervised pre-trained model. [Effects of the Invention]

[0010] The model learning device of this disclosure makes it possible to obtain a pre-trained model for constructing a highly accurate estimation model using a small amount of labeled training data. [Brief explanation of the drawing]

[0011] [Figure 1] A block diagram showing the functional configuration of the model learning device and satisfaction estimation device of Example 1. [Figure 2] A diagram showing the structure of an encoder and decoder used in model training. [Figure 3] A flowchart illustrating the speech feature reconstruction model learning operation of the model learning device in Example 1. [Figure 4] A flowchart illustrating the satisfaction estimation model learning operation of the model learning device in Example 1. [Figure 5] A flowchart illustrating the satisfaction estimation operation of the model learning device in Example 1. [Figure 6] A graph showing experimental results comparing the performance of the conventional technology and the apparatus of Example 1. [Figure 7] A graph showing experimental results comparing the performance of the conventional technology and the apparatus of Example 1. [Figure 8] A diagram illustrating an example of a computer's functional configuration. [Modes for carrying out the invention]

[0012] The embodiments of this disclosure will be described in detail below. Components having the same function will be numbered the same, and redundant explanations will be omitted.

[0013] In the embodiment, it is assumed that the utterances of each speaker included in the dialogue are recorded on separate channels for each speaker, and the channel of the target speaker is known. For example, in the case of a contact center call and the target speaker being a customer, the customer and the operator are recorded on separate channels, and it is known which channel the customer's is.

Embodiment

[0014] Referring to FIG. 1, the functional configurations of the model learning device and the satisfaction estimation device according to Embodiment 1 will be described. As shown in the figure, the model learning device 1 of the present embodiment includes a model learning unit 11 that learns a model (utterance feature reconstruction model) for reconstructing utterance features, and a model learning unit 12 that learns a satisfaction estimation model based on the utterance feature reconstruction model. Note that the model learning unit 11 and the model learning unit 12 may each be configured as an individual device. In this case, they are referred to as the model learning device 11 and the model learning device 12. The model learning unit 11 (model learning device 11) includes an audio section detection unit 111, an utterance feature extraction unit 112, an utterance feature reconstruction model learning unit 113, and an utterance feature reconstruction model storage unit 114. The model learning unit 12 (model learning device 12) includes an audio section detection unit 121, an utterance feature extraction unit 122, a satisfaction estimation model learning unit 123, and a satisfaction estimation model storage unit 124. The satisfaction estimation device 2 includes an audio section detection unit 21, an utterance feature extraction unit 22, a satisfaction estimation unit 23, and a satisfaction estimation result storage unit 24.

[0015] <Model learning unit 11 (model learning device 11)> First, as a first step, the model learning unit 11 (model learning device 11) learns a model for reconstructing utterance features using a large amount of unlabeled dialogue data. As shown in FIG. 2, the model for reconstructing utterance features is composed of a block of an encoder 51 (RNN) used for estimating utterance satisfaction in the prior art and a block of a decoder 52 that reconstructs utterance features using the output of that layer. Note that x1,...,x , ,

[0015] , t , T , ,...,x Trepresents the speech feature amounts of the 1st, …, t-th, …, T-th utterances of the target speaker, assuming that the speech feature amount of the t-th utterance is masked, and adding a “’” to the masked speech feature amount, we denote it as x’ t and denote it as x^ t represents the estimated speech feature amount of the masked target speaker.

[0016] Through the first-stage learning, the block used for estimating the speech satisfaction degree, which is the lower layer part of the model for reconstructing speech features, can obtain the tendency of the appearance probability of the sequence of speech feature amounts in the dialogue (for example, in many dialogues, it can learn that it is rare for the pitch of the customer's voice among the prosodic features to change rapidly). Hereinafter, the operations of each component in the model learning unit 11 (model learning device 11) will be described with reference to FIG. 3.

[0017] ≪Speech segment detection unit 111≫ · Input: Dialogue voice · Output: Speech sequence, speech time information The speech segment detection unit 111 acquires the dialogue voice, performs speech segment detection on each channel of the dialogue voice, and outputs a speech sequence, which is the sequence of utterances of each speaker included in the dialogue, and speech time information for each utterance (S111). The speech time information refers to the start / end time of each utterance from the start of the dialogue. In this embodiment, the speech segment detection uses a method based on a power threshold process, but other speech segment detection methods such as a method based on the likelihood ratio of a speech / non-speech model may also be used.

[0018] ≪Speech feature extraction unit 112≫ · Input: Speech sequence, speech time information · Output: Speech feature sequence The speech feature extraction unit 112 acquires the speech sequence and speech time information, extracts the speech features corresponding to each utterance of the target speaker, and outputs a speech feature sequence, which is a sequence of speech features (S112). For example, the speech features can be any one or more of prosodic features, dialogue features, and linguistic features.

[0019] As prosodic features, at least one of the following will be used: the mean, standard deviation, maximum, and minimum values ​​of the fundamental frequency and power during the speaker's utterance; the speaker's speech rate during the utterance; and the duration of the final phoneme during the speaker's utterance. Here, the fundamental frequency and power will be determined frame by frame by dividing the utterance into frames. When using speech rate and the duration of the final phoneme, speech recognition will be used to estimate the phoneme sequence during the utterance.

[0020] As dialogue features, at least one of the following is used: the time since the target speaker's previous utterance, the time between the previous non-target speaker's utterance and the target speaker's utterance, the time between the target speaker's utterance and the next non-target speaker's utterance, the length of the target speaker's utterance, the lengths of the preceding and succeeding non-target speakers' utterances, the number of interjections made by the target speaker during the preceding and succeeding non-target speakers' utterances, and the number of interjections made by the non-target speakers during the target speaker's utterance.

[0021] As linguistic features, at least one of the following will be used: the number of words in the target speaker's utterance, the number of fillers in the target speaker's utterance, and the number of occurrences of words of gratitude in the target speaker's utterance. When using linguistic features, speech recognition will be used to estimate the words appearing in the utterance, and the results will be used. In addition, words of gratitude will be selected manually, for example, the number of occurrences of "thank you" or "thanks" will be determined.

[0022] ≪Speech Feature Reconstruction Model Learning Unit 113≫ • Input: Utterance feature sequence • Output: Speech feature reconstruction model (unsupervised pre-trained model) The speech feature reconstruction model learning unit 113 acquires speech feature sequences, randomly selects a portion of the speech feature sequences corresponding to each utterance of the target speaker and replaces them with predetermined masking information to mask them, and learns a speech feature reconstruction model, which is a neural network model that estimates the speech features of the masked speech feature sequences. The learned speech feature reconstruction model is then output as an unsupervised pre-trained model (S113).

[0023] Masking refers to the process of replacing a feature with a vector that has the same dimensionality but holds different values, for example, by making the feature a zero vector. The parameters of the model that reconstructs the speech features are updated using these masked speech features as input to estimate the speech features in the masked areas.

[0024] The speech feature reconstruction model learning unit 113 preferably randomly selects speech feature sequences and randomly masks some of the selected speech feature quantities. For example, the speech feature reconstruction model learning unit 113 randomly selects the locations to be masked, then replaces 80% of the locations with zero vectors, 10% with speech features from other random locations in the dialogue, and leaves 10% unreplaced. Furthermore, it is preferable that the locations to be masked represent a maximum of 20% of the speech features in the entire dialogue. The proportion of locations to be masked and the masking method may be changed or deleted; for example, masking may be performed by replacing with the average value of speech features in the entire dialogue.

[0025] For example, a Long Short-Term Memory Recurrent Neural Network (LSTM-RNN) and a decoder based on fully connected layers can be used as a speech feature reconstruction model. However, layers of neural networks other than fully connected layers and LSTM-RNN may also be used; for example, a Gated Recurrent Unit may be used instead of an LSTM-RNN.

[0026] For model training, we use Back Propagation Through Time, an existing neural network training method. The L1 norm of the features is used as the loss function, but another distance measure (e.g., the L2 norm) may also be used.

[0027] ≪Speech Feature Reconstruction Model Memory Unit 114≫ • Input: Speech feature reconstruction model • Output: Speech feature reconstruction model The speech feature reconstruction model storage unit 114 stores the speech feature reconstruction model learned and output in step S113, and outputs the stored speech feature reconstruction model in response to a request from the model learning unit 12 (model learning device 12) (S114).

[0028] <Model Learning Unit 12 (Model Learning Device 12)> In the second stage, the model learning unit 12 (model learning device 12) uses a small amount of labeled training data to train an estimation model (satisfaction estimation model) that estimates utterance satisfaction and dialogue satisfaction. As shown in Figure 2, the satisfaction estimation model consists of a block of encoder 61 (RNN) used for estimating utterance satisfaction, a block of decoder 62 that uses the output of the encoder 61 layer, a block of encoder 63 (RNN) used for estimating dialogue satisfaction, and a block of decoder 64 that uses the output of the encoder 63 layer. Note that u1, ..., u t ,…,u T d represents the utterance satisfaction of the 1st, ..., t, ..., Tth utterances of the target speaker, and d represents the dialogue satisfaction.

[0029] The learned parameters of encoder 51 obtained in the first stage of learning are used as the initial model parameters of encoder 61, which is used to estimate speech satisfaction in the estimation model. The parameter update is performed using the same method as conventional methods, but with a reduced learning rate (for example, 1 / 10 of that of conventional techniques that do not perform pre-training). This allows the estimation model to learn to estimate speech satisfaction and dialogue satisfaction while considering the likelihood of occurrence of sequences of speech features in dialogue. As a result, an estimation model capable of estimating speech satisfaction and dialogue satisfaction with high accuracy can be obtained even when using a small amount of labeled training data (for example, if the pitch of the target speaker's voice changes abruptly in an utterance where the speech satisfaction was "satisfied," it may be difficult to link the abrupt change in the target speaker's voice pitch to the "satisfied" estimation result with only a small amount of training data. However, by performing the first stage of learning, the model acquires the understanding that abrupt changes are rare = characteristic, making it easier to link them to the "satisfied" estimation result).

[0030] This approach of using unlabeled data for pre-training is called unsupervised pre-training, and its effectiveness has been confirmed in the fields of natural language processing and image processing (see Non-Patent Documents 1 and 2). However, there are no examples of using unsupervised pre-training to estimate the utterance satisfaction and dialogue satisfaction of target speakers in dialogues, nor are there any examples of using unsupervised pre-training for heuristic features such as prosodic features, dialogue features, and linguistic features.

[0031] (Reference Non-Patent Document 1: Ting Chen and Simon Kornblith and Mohammad Norouzi and Geoffrey Hinton, “A Simple Framework for Contrastive Learning of Visual Representations,” Proc. ICML, pp. 1597-1607, 2020.) (Reference Non-Patent Document 2: Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova, “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding,” Proc. of NAACL-HLT, pp. 4171-4186, 2019.) The operation of each component within the model learning unit 12 (model learning device 12) will be explained below with reference to Figure 4.

[0032] ≪Audio section detection unit 121≫ • Input: Conversational voice • Output: Utterance sequence, utterance time information The voice segment detection unit 121 performs the same processing as the voice segment detection unit 111 (S121). However, the dialogue audio for which voice segments are detected is data labeled with speech satisfaction and dialogue satisfaction.

[0033] <<Speech Feature Extraction Unit 122>> • Input: Utterance sequence, utterance time information • Output: Utterance feature sequence The speech feature extraction unit 122 performs the same processing as the speech feature extraction unit 112 based on the speech sequence and speech time information output in step S121 (S122).

[0034] ≪Satisfaction Estimation Model Learning Unit 123≫ • Input: Utterance feature sequence, Utterance satisfaction label, Dialogue satisfaction label, Utterance feature reconstruction model Output: Satisfaction estimation model The satisfaction estimation model learning unit 123 uses the parameters of the speech feature reconstruction model as initial values ​​for the model parameters, and uses the speech feature sequence and the corresponding speech satisfaction labels and dialogue satisfaction labels as training data to perform supervised learning of a satisfaction estimation model, which is a model that estimates speech satisfaction and dialogue satisfaction (S123).

[0035] In this embodiment, an LSTM-RNN or fully connected layer is used for the satisfaction estimation model. In this case, since the satisfaction estimation model is trained using the parameters of the speech feature reconstruction model as initial values, the speech satisfaction estimation portion of the speech feature reconstruction model and the speech satisfaction and dialogue satisfaction estimation models will use an LSTM-RNN with the same number of hidden layers and units.

[0036] The model learning method will be the same as conventional methods. That is, the model parameters will be updated by backpropagating the loss error obtained by the weighted sum of the estimated loss errors of utterance satisfaction and dialogue satisfaction. However, to avoid significantly changing the parameters learned by the utterance feature reconstruction model, the learning rate will be kept small (for example, α = 0.0001 when using Adam as the optimization method) when updating the parameters.

[0037] <<Satisfaction Estimation Model Memory Unit 124>> • Input: Satisfaction estimation model Output: Satisfaction estimation model The satisfaction estimation model storage unit 114 stores the satisfaction estimation model learned and output in step S123, and outputs the stored satisfaction estimation model in response to a request from the satisfaction estimation device 2 (S124).

[0038] <Satisfaction Estimation Device 2> The satisfaction estimation device 2 estimates utterance satisfaction and dialogue satisfaction based on the satisfaction estimation model learned in the second stage. The operation of each component within the satisfaction estimation device 2 will be explained below with reference to Figure 5.

[0039] ≪Audio section detection unit 21≫ • Input: Conversational voice • Output: Utterance sequence, utterance time information The voice segment detection unit 21 performs the same processing as the voice segment detection units 111 and 121 (S21). However, the dialogue audio for which the voice segment is detected is the dialogue audio targeted for satisfaction estimation.

[0040] ≪Speech Feature Extraction Unit 22≫ • Input: Utterance sequence, utterance time information • Output: Utterance feature sequence The speech feature extraction unit 22 performs the same processing as the speech feature extraction unit 112 and the speech feature extraction unit 122 based on the speech sequence and speech time information output in step S21 (S22).

[0041] ≪Satisfaction Estimation Section 23≫ • Input: Speech feature sequence, speech satisfaction level, satisfaction estimation model (If the model is stored in the satisfaction estimation device 2, the model input is only required the first time.) Output: Estimated results series of speech satisfaction, estimated results of dialogue satisfaction The satisfaction estimation unit 23 estimates utterance satisfaction and dialogue satisfaction based on the satisfaction estimation model learned in the second stage, obtains the estimated sequence of utterance satisfaction and the estimated result of dialogue satisfaction, and outputs them (S23). The satisfaction estimation unit 23 inputs the sequence of utterance features into the satisfaction estimation model and performs forward propagation to simultaneously obtain the estimated sequence of utterance satisfaction and the estimated result of dialogue satisfaction.

[0042] <<Satisfaction Estimation Result Storage Unit 24>> • Input: Estimated results series of speech satisfaction, estimated results of dialogue satisfaction Output: Estimated results series of speech satisfaction, estimated results of dialogue satisfaction The satisfaction estimation result storage unit 24 stores the estimated speech satisfaction result sequence and the estimated dialogue satisfaction result output in step S23, and outputs the stored estimated speech satisfaction result sequence and the estimated dialogue satisfaction result upon request from any device (S24).

[0043] In the above disclosure, the method disclosed in Patent Document 1 can be applied, except that the parameters of the encoder portion of the model are changed to those of a pre-trained model that has been trained with labels.

[0044] However, any method for estimating speech satisfaction or dialogue satisfaction from speech feature sequences is acceptable, and the specific method is not limited to Patent Document 1.

[0045] For example, even with models other than those that estimate utterance satisfaction and dialogue satisfaction in two stages, if the input is the same as that of Patent Document 1 and similar processing is performed, the initial parameters from the pre-trained model described above will be effective.

[0046] In particular, the initial parameters from the pre-trained model described above are effective for inference models that fall under the following two categories:

[0047] • Extracting features from each utterance of the target speaker. • The model must perform inferences based on either the target speaker's utterance sequence or the entire call.

[0048] Another example where initial parameters from a pre-trained model are effective is, for example, cold anger detection.

[0049] <Evaluation experiment> Figure 6 compares the relationship between the error rate of estimated dialogue satisfaction by the model trained by the model learning device 1 and the amount of labeled training data, with the relationship between the error rate of estimated dialogue satisfaction by the conventional technology model (a model trained only with labeled training data) and the amount of labeled training data. From this figure, it can be seen that the model learning device 1 of this embodiment achieves the same level of accuracy as the conventional technology even when the amount of labeled training data is reduced by 50%.

[0050] Figure 7 compares the estimated error rates for speech satisfaction and dialogue satisfaction obtained by the model trained by Model Learning Device 1, using the same amount of labeled training data as the conventional technology, with the estimated error rates for speech satisfaction and dialogue satisfaction obtained by the conventional technology model (a model trained only on labeled training data). From this figure, it can be seen that when the amount of labeled training data is the same as the conventional technology, Model Learning Device 1 is able to reduce the estimated error rates for speech satisfaction / dialogue satisfaction by more than 10% compared to the conventional technology.

[0051] <Effects> The model learning device 1 and satisfaction estimation device 2 of Example 1 are characterized by performing pre-training of the estimation model using unlabeled dialogues. This feature allows for obtaining a highly accurate estimation model even when using a small amount of labeled training data. This makes it possible to provide applications that require estimation of utterance satisfaction and dialogue satisfaction (e.g., automation of operator evaluation in call centers) inexpensively and with high reliability.

[0052] The model learning device 1 and satisfaction estimation device 2 of Example 1 have an additional element compared to conventional systems such as those in Patent Document 1, which uses a pre-trained model. This additional element enumerates specific methods for reducing the amount of labeled training data required or reducing the estimation error rate with the same amount of labeled training data compared to conventional systems, thereby providing a reduction in the amount of computation required by the computer and an improvement in the accuracy of computer estimation.

[0053] <Note> The device of this disclosure, for example, as a single hardware entity, has an input section to which a keyboard or the like can be connected, an output section to which a liquid crystal display or the like can be connected, a communication section to which a communication device (e.g., a communication cable) that can communicate with the outside of the hardware entity can be connected, a CPU (Central Processing Unit, which may include cache memory and registers), RAM or ROM as memory, an external storage device such as a hard disk, and a bus that connects these input section, output section, communication section, CPU, RAM, ROM, and external storage device to enable data exchange between them. Furthermore, if necessary, the hardware entity may be equipped with a device (drive) that can read and write recording media such as CD-ROMs. An example of a physical entity equipped with such hardware resources is a general-purpose computer.

[0054] The external storage device of the hardware entity stores the programs necessary to realize the above-mentioned functions, as well as the data required for processing these programs (this is not limited to external storage; for example, programs may be stored in ROM, a read-only storage device). Furthermore, data obtained through the processing of these programs is appropriately stored in RAM or other external storage devices.

[0055] In hardware entities, each program stored in an external storage device (or ROM, etc.) and the data necessary for processing each program are loaded into memory as needed, and interpreted, executed, and processed by the CPU as appropriate. As a result, the CPU realizes the predetermined functions (each of the configuration requirements expressed as ...part, ...means, etc. above).

[0056] This disclosure is not limited to the embodiments described above, and may be modified as appropriate without departing from the spirit of this disclosure. Furthermore, the processes described in the embodiments above may not only be executed in chronological order according to the order described, but may also be executed in parallel or individually as needed, depending on the processing capacity of the device performing the processes.

[0057] As described above, when the processing functions of the hardware entity (device of this disclosure) described in the above embodiment are implemented by a computer, the processing content of the functions that the hardware entity should have is described by a program. Then, by executing this program on the computer, the processing functions of the hardware entity are implemented on the computer.

[0058] The various processes described above can be carried out by loading a program that executes each step of the above method into the recording unit 10020 of the computer 10000 shown in Figure 8, and then causing the control unit 10010, input unit 10030, output unit 10040, etc. to operate.

[0059] The program describing this processing can be recorded on a computer-readable recording medium. Any computer-readable recording medium can be used, such as a magnetic recording device, optical disc, magneto-optical recording medium, or semiconductor memory. Specifically, examples include hard disk drives, flexible disks, and magnetic tapes as magnetic recording devices; DVDs (Digital Versatile Discs), DVD-RAMs (Random Access Memory), CD-ROMs (Compact Disc Read Only Memory), and CD-Rs (Recordable) / RWs (ReWritable) as optical discs; MOs (Magneto-Optical Discs) as magneto-optical recording media; and EEP-ROMs (Electrically Erasable and Programmable-Read Only Memory) as semiconductor memory.

[0060] Furthermore, this program may be distributed, for example, by selling, transferring, or lending portable recording media such as DVDs or CD-ROMs on which the program is recorded. Alternatively, the program may be stored in the storage device of a server computer and distributed by transferring the program from the server computer to other computers via a network.

[0061] A computer executing such a program may, for example, first store the program recorded on a portable storage medium or a program transferred from a server computer in its own storage device. Then, when processing is to be executed, the computer reads the program stored on its own storage medium and executes the processing according to the read program. Alternatively, the computer may directly read the program from the portable storage medium and execute the processing according to that program, or it may sequentially execute the processing according to the received program each time a program is transferred to it from a server computer. Furthermore, the above processing may be executed by a so-called ASP (Application Service Provider) type service, where the server computer does not transfer programs to this computer, but the processing function is realized only by execution instructions and result acquisition. In this form, the program includes information used for processing by an electronic computer that is equivalent to a program (data that is not a direct instruction to the computer but has the property of defining the processing of the computer, etc.).

[0062] Furthermore, in this configuration, a hardware entity is constructed by executing a predetermined program on a computer, but at least a part of these processes may be implemented in hardware.

[0063] The following additional information is disclosed regarding the embodiments described above.

[0064] (Additional note 1) Memory and At least one processor connected to the memory, Includes, The aforementioned processor, The process involves randomly selecting a portion of the speech feature sequence, which is a sequence of speech features corresponding to each utterance of a target speaker, and replacing it with predetermined masking information to mask it. A speech feature reconstruction model, which is a neural network model that estimates the speech features of the masked speech feature sequence, is then trained, and the trained speech feature reconstruction model is output as an unsupervised pre-trained model. Model learning device.

[0065] (Additional note 2) A non-temporary storage medium that stores a program executable by a computer to perform a model learning process, The aforementioned model learning process is, The process involves randomly selecting a portion of the speech feature sequence, which is a sequence of speech features corresponding to each utterance of a target speaker, and replacing it with predetermined masking information to mask it. A speech feature reconstruction model, which is a neural network model that estimates the speech features of the masked speech feature sequence, is then trained, and the trained speech feature reconstruction model is output as an unsupervised pre-trained model. Non-transitory storage medium.

[0066] (Additional note 3) The model learning device described in Appendix 1, The aforementioned processor, Using the parameters of the aforementioned speech feature reconstruction model as initial values ​​for the model parameters, and using the speech feature sequence and corresponding speech satisfaction labels and dialogue satisfaction labels as training data, a satisfaction estimation model, which estimates speech satisfaction and dialogue satisfaction, is trained using supervised learning. Model learning device.

[0067] (Additional note 4) A non-temporary storage medium as described in Appendix 2, The aforementioned model learning process is, Using the parameters of the aforementioned speech feature reconstruction model as initial values ​​for the model parameters, and using the speech feature sequence and corresponding speech satisfaction labels and dialogue satisfaction labels as training data, a satisfaction estimation model, which estimates speech satisfaction and dialogue satisfaction, is trained using supervised learning. Non-transitory storage medium.

[0068] (Additional note 5) The model learning device described in Appendix 1, The aforementioned speech features are one of the following: prosodic features, dialogue features, or linguistic features. Model learning device.

[0069] (Additional note 6) A non-temporary storage medium as described in Appendix 2, The aforementioned speech features are one of the following: prosodic features, dialogue features, or linguistic features. Non-transitory storage medium.

[0070] (Additional note 7) Memory and At least one processor connected to the memory, Includes, The aforementioned processor, The process involves randomly selecting a portion of the speech feature sequence, which is a sequence of speech features corresponding to each utterance of the target speaker, and replacing it with predetermined masking information to mask it. The parameters of the speech feature reconstruction model, which is a neural network model that estimates the speech features of the masked speech feature sequence, are used as the initial values ​​for the model parameters. Based on a satisfaction estimation model trained using the speech feature sequence and the corresponding speech satisfaction labels and dialogue satisfaction labels as training data, the speech satisfaction and dialogue satisfaction corresponding to the target speaker's utterances are estimated. Satisfaction estimation device.

[0071] (Additional note 8) A non-temporary storage medium that stores a program executable by a computer to perform satisfaction estimation processing, The satisfaction estimation process described above is: The process involves randomly selecting a portion of the speech feature sequence, which is a sequence of speech features corresponding to each utterance of the target speaker, and replacing it with predetermined masking information to mask it. The parameters of the speech feature reconstruction model, which is a neural network model that estimates the speech features of the masked speech feature sequence, are used as the initial values ​​for the model parameters. Based on a satisfaction estimation model trained using the speech feature sequence and the corresponding speech satisfaction labels and dialogue satisfaction labels as training data, the speech satisfaction and dialogue satisfaction corresponding to the target speaker's utterances are estimated. Non-transitory storage medium.

Claims

1. The system includes a speech feature reconstruction model learning unit that randomly selects a portion of a sequence of speech features, which are speech features corresponding to each utterance of a target speaker and are either prosodic features, dialogue features, or linguistic features, and replaces them with predetermined masking information to mask them; trains a speech feature reconstruction model, which is a neural network model that estimates the speech features of the masked sequence of speech features; and outputs the trained speech feature reconstruction model as an unsupervised pre-trained model. Model learning device.

2. A model learning device according to claim 1, The unit includes a satisfaction estimation model learning unit that supervises learning of a satisfaction estimation model, which is a model that estimates utterance satisfaction and dialogue satisfaction, using the parameters of the aforementioned utterance feature reconstruction model as initial values ​​for the model parameters, and the utterance feature sequence and corresponding utterance satisfaction labels and dialogue satisfaction labels as training data. Model learning device.

3. A model learning method performed by a model learning device, The process includes the steps of randomly selecting a portion of a sequence of speech features, which are speech features corresponding to each utterance of a target speaker and are either prosodic features, dialogue features, or linguistic features, and replacing them with predetermined masking information to mask them; training a speech feature reconstruction model, which is a neural network model that estimates the speech features of the masked sequence of speech features; and outputting the trained speech feature reconstruction model as an unsupervised pre-trained model. Model learning methods.

4. A program that causes a computer to function as a model learning device as described in claim 1.

Images