JP6864322B2 - Voice processing device, voice processing program and voice processing method - Google Patents

Description

The present invention relates to a voice processing device, a voice processing program and a voice processing method, and more particularly to a voice processing device, a voice processing program and a voice processing method capable of speaking with emotional expression.

An example of a speech synthesizer that is the background of the present invention is disclosed in Patent Document 1. The technique of Patent Document 1 can provide a utterance style conversion system based on a neural network, which can transform the utterance style and speaker character of a certain speaker based on an arbitrary control law.

Japanese Patent Application Laid-Open No. 2017-32839 [G10L 13/06]

In the technique of Patent Document 1, various speaker characteristics are reproduced by using numerical data representing the characteristics of the speaker in the input unit of the neural network acoustic model for speech synthesis. Since the learning is not performed in consideration of the internal behavior when the above is changed, there is a problem that there is no guarantee that the desired speaker character can be reproduced.

Therefore, a main object of the present invention is to provide a novel voice processing device, a voice processing program and a voice processing method.

Another object of the present invention is to provide a speech processing device, a speech processing program, and a speech processing method capable of optimally setting a change in the speech style of the output speech with respect to the input speech.

Another object of the present invention is to provide a voice processing device, a voice processing program, and a voice processing method that can easily realize a desired speaker character.

A first aspect of the present invention is undisturbed speech style voice and at least one speech corpus including the feature quantity data based on the audio data calm than the speech style voice, calm the speech style voice and at least one undisturbed other speech style voice control parameter generation unit for setting a control parameter including a first mixing ratio is mixing ratio, learning data and calm speech style of the input data including control parameters, including the feature amount data and the first mixing ratio of undisturbed speech style voice A training data generator that generates training data for output data including a conversion filter calculated by a control parameter that includes a first mixing ratio from voice and at least one non-quiet speech style voice feature data, and input data. It is a voice processing device including a learning unit that learns an input-output conversion model based on training data of training data and output data.

In the first invention, a voice processing device (10: a reference reference numeral indicating a corresponding portion in an embodiment; the same applies hereinafter) is used in a voice corpus storage unit (20, 101) other than a calm voice and at least one calm. A voice corpus including feature data based on the voice data of the speech style voice is stored. The control parameter generation unit (S5, 107) is, for example, a first mixing ratio of the calm voice and at least one non-quiet speech style voice within the definition area of the control law set in the control law storage unit (109). Generate control parameters including mixing ratio. The learning data generation unit (S7, 105) includes learning data of input data including feature data of calm utterance style voice and control parameters including a first mixing ratio, calm utterance style voice, and at least one non-quiet utterance. From the feature amount data of the style voice, the training data of the output data including the conversion filter calculated by the control parameter including the first mixing ratio is generated. Then, the learning unit (S9-S15, 111) learns the input-output conversion model based on the learning data of the input data and the learning data of the output data.

According to the first invention, when learning based on the voice data of emotional speech (speech style speech other than calm) and calm style speech, the calm speech style speech and at least one non-calm speech style speech are mixed. Since the training data is created by changing the ratio of the speech, it is possible to optimally set the change of the utterance style in the output speech with respect to the input speech in the conversion model.

The second invention is dependent on the first invention, the voice corpus contains feature data based on the voice data of a calm voice and at least one non-quiet speech style voice for each of the different speakers, control parameters. Includes a second mixing ratio, which is a mixing ratio of multiple speakers to the same speech style voice , and the learning data generator follows control data including the first mixing ratio and the second mixing ratio of the speaking style and the speaker, respectively. It is a voice processing device that generates training data.

In the second invention, the voice corpus stored in the voice corpus storage units (20, 101) provides features of voice data of calm voice and at least one non-quiet speech style voice for each of a plurality of different speakers. The control parameter generated by the control parameter generation unit (S5, 107) includes a second mixing ratio, which is a mixing ratio of the same speech style voice of a plurality of speakers. Therefore, the learning data generating unit generates the learning data according to the control data including the respective mixture ratio of the speech style and a speaker.

According to the second invention, if each emotional voice ( speech style voice other than calm) of a plurality of speakers is learned based on the voice data of the calm utterance style voice, the output voice of the desired speaker character can be obtained. It can be easily output.

The third invention is a conversion filter prediction unit that predicts a conversion filter using the input-output conversion model learned by the speech processing apparatus of the first invention or the second invention, and a conversion filter that converts an input waveform. A voice processing device including a waveform conversion unit.

In the third invention, the conversion filter prediction unit (S27, 209) predicts the conversion filter using the input-output conversion model learned by the speech processing apparatus of the first invention or the second invention. The waveform conversion unit (S29, 211) converts the input waveform using the conversion filter.

According to the third invention, a desired speech style and / or speaker-like voice waveform can be output from the waveform conversion unit.

A fourth invention is a computer processor that is executed by a computer and uses a voice corpus that includes feature data based on calm speech style voice and at least one non-quiet speech style voice voice data. A control parameter generator that sets a control parameter including a first mixing ratio, which is a mixing ratio of a calm speech style voice and at least one non- quiet speech style voice, a feature amount data of a calm speech style voice, and a first. training data and serenity of speech style voice input data including control parameters, including the mixing ratio with a conversion filter which is calculated by the control parameters including a first mixing ratio of the feature data of the at least one undisturbed other speech style voice It is a voice processing program that functions as a learning data generation unit that generates training data of the including output data, and a learning unit that learns an input-output conversion model based on the training data of the input data and the training data of the output data.

A fifth invention is a voice processing method executed by a computer and using a voice corpus containing feature data based on voice data of a calm speech style voice and at least one non-quiet speech style voice, wherein the calm speech is performed. Control parameter generation step to set the control parameter including the first mixing ratio which is the mixing ratio of the style voice and at least one non-quiet speech style voice, the control including the feature amount data of the calm speech style voice and the first mixing ratio. Training of input data including parameters and learning of output data including conversion filters calculated by control parameters including a first mixing ratio from feature data of calm speech style speech and at least one non-quiet speech style speech. It is a voice processing method including a learning data generation step for generating data and a learning step for learning an input-output conversion model based on training data of input data and training data of output data.

The same effect as that of the first invention can be expected by the fourth invention or the fifth invention.

According to the present invention, different emotional voices (speech style voices) of the same speaker are recorded separately from the calm voices, and the ratio of mixing the speech style voices when learning based on each voice data is changed. Since the training data is created by the above, the change of the utterance style in the output speech with respect to the input speech can be optimally set in the conversion model.

Further, if the emotional voice (speech style voice) of each of the plurality of speakers is learned based on the voice data of the calm voice, the output voice of a desired speaker can be easily output.

The above-mentioned objectives, other objectives, features and advantages of the present invention will become more apparent from the detailed description of the following examples made with reference to the drawings.

FIG. 1 is a block diagram showing an example of an electrical configuration of an audio processing device according to an embodiment of the present invention. FIG. 2 is an illustrated diagram showing an example of the voice corpus of the embodiment of FIG. FIG. 3 is an illustrated diagram showing an example of a memory map of the memory of the embodiment of FIG. FIG. 4 is a flow chart showing an example of the learning operation of the embodiment of FIG. FIG. 5 is an illustrated diagram showing an example of learning parameters applied to the embodiment of FIG. FIG. 6 is a functional block diagram corresponding to the flow diagram of FIG. 4 in the embodiment of FIG. FIG. 7 is a graph showing an example of conversion characteristics in the embodiment of FIG. 1 in comparison with the conventional one. FIG. 8 is a flow chart showing an example of the output operation in the embodiment of FIG. FIG. 9 is a functional block diagram corresponding to the flow diagram of FIG. 8 in the embodiment of FIG.

With reference to FIG. 1, the voice processing device 10 of this embodiment is basically a general-purpose computer, includes a CPU (processor) 12, the CPU 12 is connected to the communication module 16 through the bus 14, and therefore the CPU 12 , When necessary, is communicably connected to a network (not shown) via the communication module 16.

The CPU 12 can also access the memory 18 and the HDD 20 through the bus 14, and constructs a conversion model according to voice processing, for example, a deep neural network (DNN) according to the programs and data (described later) set in the HDD 20 and the memory 18. , And / or use such a conversion model to convert the input audio waveform to the output audio waveform. That is, the voice output mechanism of the voice processing device 10 of this embodiment is a waveform connection type voice synthesis, and generates an output voice waveform while connecting voice waveforms recorded according to the utterance style, for example, for each phoneme.

The output voice waveform is given to the speaker 22 as an analog voice signal from the CPU 12 via an appropriate interface (not shown). Therefore, the output voice obtained by converting the input voice by the conversion model is output from the speaker 22.

Further, the audio processing device 10 includes a display 23 such as an LCD.

The HDD 20 in FIG. 1 is used as a database, and here functions as a voice corpus storage unit. However, the HDD is only given as an example of a large-capacity storage device for convenience, and another semiconductor storage device such as USB or an optical storage device may be used.

The voice corpus storage unit is composed of a speech style voice and a calm voice having the same speech content. Speaking style voice refers to voice that has a voice impression such as "bright", "dark", and "cute", and a voice quality that expresses human emotions such as "surprise", "anger", and "joy". In this embodiment, three of the utterance styles, "angry", "sad" and "joy", were adopted. In the experiments of the inventors, the voices of four speakers were recorded to create a corpus as shown in FIG. However, if there is no same utterance among the utterance style voice and the calm voice, it is created from the synthetic voice.

The parameters to be generated include the spectral envelope, the fundamental frequency, the asynchronous index, the residual vector, the phase feature, and the parameters related to the delta component representing their time variation. And examples of parameters related to spectrum envelope include FFT (Fast Fourier Transform) spectrum, cepstrum, melkeptrum, line spectrum pair, mel line spectrum pair, mel generalized spectrum, coding component of self-encoder and the like.

In the embodiment, a cepstrum (a cepstrum (Fourier-transformed voice spectrum) in consideration of human auditory characteristics) is adopted. FIG. 2 shows a corpus created by dividing the “angry” utterance-style voice among the calm voice and the utterance-style voice into, for example, every 5 ms (1 frame) and FFT). That is, data indicating the feature amount included in one frame of the input voice is generated as a corpus. However, the parameters of the calm voice of the same speaker and each utterance style voice must be consistent in the time axis direction.

In FIG. 2, a vertically long strip illustrates one feature data, and matching is determined between the feature data of the calm voice and the feature data of each utterance style voice, and the feature data of the utterance style voice is the same as the feature data of the calm voice. Is omitted. In this embodiment, as will be described later, since the parameter indicating the speaker ID is calculated from the characteristic data of the quiet voice, the voice characteristic data of the quiet voice is not omitted.

In the embodiment, a voice corpus in which the calm voices of the plurality of speakers and the voices of the plurality of utterance styles are recorded is prepared in advance, and the model learning process shown in FIG. 4 is executed and converted ( (Input-output conversion) model is constructed and stored in a model storage unit such as HDD 20 shown in FIG.

In the embodiment, a voice corpus was created using the utterances of each of a plurality of (for example, four) speakers, each of which is a calm voice and three utterance style voices. However, the number of speakers may be one, and the number of speech-style voices may be one or more, or more.

As shown in FIG. 3, the memory 18 of FIG. 1 is provided with a program storage unit 24 and a data storage unit 26.

The learning program 24a for the learning process shown in FIG. 4 and the conversion program 24b for the output conversion process shown in FIG. 8 are preset in the program storage unit 24. However, when the voice processing device 10 of the embodiment is used only for model construction, only the learning program 24a needs to be set, and the voice processing device 10 is used only for output voice conversion. If so, only the conversion program 24b needs to be set.

Although will be described later, the data storage unit 26 has a learning parameter storage area 26a, a control parameter storage area 26b, and a control law storage area for storing learning parameters, control parameters, control rules, and learning data used in the learning process, respectively. 26c and the learning data storage area 26d are formed.

Further, an input audio waveform storage area 26e, an input parameter storage area 26f, and a model storage area 26g for storing the input audio waveform, the input parameter, and the conversion model used for the conversion process are formed.

Further, the data storage area 26 includes an area such as a counter (not shown) for counting the number of times i of learning to be repeated, which will be described later.

With reference to FIG. 4, in the first step S1 of the learning process, the learning parameters as shown in FIG. 5 are generated. That is, the step S1 executed by the CPU 12 (FIG. 1) functions as a learning parameter generation unit.

The learning parameters here include input voice parameters and output voice parameters. Here, the input voice parameter is the data of the feature amount of the voice shown in FIG. 2, and is a voice corpus. The speaking style shows anger, sadness, and joy. The learning parameter generation unit further generates a parameter representing the ID (identification symbol) of the input speaker (four people in the embodiment) and a parameter representing the ID of the utterance style of the output voice. The output voice parameter is a parameter of the utterance style of the output voice, and is a parameter indicating the utterance style voice of the output voice.

As the parameter representing the ID of the input speaker, the parameter calculated from the voice corpus of the calm voice of the speaker is adopted. However, for the calculation, for example, one hot vector, i-vector component, principal component score (similarity, calculation index) of principal component analysis, output coefficient of the encoder of autoencoder, etc. are used. However, in the examples, the principal component score of the principal component analysis is used.

As the utterance style parameter of the output voice, the one-hot vector or the principal component score of the principal component analysis can be assumed, but in the embodiment, the one-hot vector is used.

In the next step S3, the counter i (not shown) formed in the data storage area 26 of the memory 18 is initialized (i = 1). This is to count the number of times the learning process after step S5 is repeated.

In the next step S5 of FIG. 4, a control parameter is generated. That is, the step S5 executed by the CPU 1 functions as a control parameter generation unit. The control parameter is a parameter of a mixing ratio (first mixing ratio) indicating the ratio of mixing the above three utterance style voices, which is one of the features of this embodiment. In this step S5, control parameter generation conditions are set (stored) as one setting file each time a conversion model is created.

As a condition for generating control parameters, in addition to the number of parameters to be generated at one time, there is a method (fixed value or random number) for generating each control parameter. As an example, the number of speakers generated at one time is the number of speakers in the above-mentioned voice corpus (“4” in the example) + the number of utterance styles (“3” in the example).

The control parameter generation method (control rule) indicates whether the above-mentioned mixing ratio is set with a fixed value or is determined by using a random number. For fixed values, specify each parameter with a fixed value. When specified by a random number, the random number for the speaker is a real number in which the total of "0" to "1" is "1", and the random number for the utterance style is a real number from "0" to "1". Since the former sets the ratio of each speaker (second mixing ratio), the total needs to be "1", but in the latter case (first mixing ratio), such regulation is not necessary. However, the ratio of speakers may be set to a fixed value, and only the utterance style may be changed by random numbers. Such control rules for generating control parameters are preset by the designer of the voice processing device 10.

Then, this control parameter generation unit generates a control parameter indicating the mixing ratio based on such a generation condition, and stores (sets) it in the control parameter storage unit. However, the generated control parameter is stored in association with the control parameter generation condition setting file.

In the following step S7, the CPU 12 generates learning data based on the control parameters generated in step S5 and the control rule read from the control rule storage area 26c of the data storage area 26. That is, the step S7 executed by the CPU 12 functions as a learning data generation unit. In this step S7, the learning parameters shown in FIG. 5 are transformed according to the control law and the control parameter values to generate learning data.

In the learning data generation unit, that is, step S7, as described above, the learning data is generated according to the control parameters and the control law, but here, as an example, the control law is a linear combination (control linearly). The process of generating the training data in the case will be described. However, it is assumed that all the utterances are the same.

The number 1 is an input voice parameter, for example, 25-dimensional data.

Here, v _n is a control parameter (second mixing ratio) for the nth speaker, and x _t ⁽ⁿ⁾ indicates the calm voice of the nth speaker.

Equation 2 is a parameter of the output voice, for example, 25-dimensional data.

Here, u _m is a control parameter (first mixing ratio) for the m-th utterance style, and y _t ^(m) is a parameter indicating the m-th utterance style.

Equation 3 is a parameter indicating the ID of the utterance style of the output voice, and is, for example, a three-dimensional vector.

Here, e _n is a parameter indicating an ID of the n-th speech style.

The number 4 is a parameter representing the ID of the nth speaker, and is, for example, a three-dimensional vector.

Here, S _n is a parameter representing the ID of the nth speaker.

Of the data obtained by the processes of Equations 1 to 4, the 31-dimensional vector represented by Equation 5 is used as the training data of the input data, and Equation 6 is used as the training data of the output data.

Where T represents transpose.

Here, dt is the learning data of the output data.

In this way, in step S7, the training data is generated, and in the next step S9, the CPU 12 learns the conversion model. That is, step S9 executed by the CPU 12 functions as a model learning unit.

However, as the training data of the output data, VQ (d _t ) obtained by quantizing this may be used _{instead of the d t of the equation 6.}

The model constructed here assumes a deep neural network (DNN). Further, as the type of network, a feedforward network, a convolutional network, a recurrent network, etc. are assumed, but in the embodiment, the feedforward network is adopted.

Further, the model learning method in step S9 can utilize the stochastic gradient descent method generally used in DNN, the normalization and learning rate scheduling method associated therewith, and hostile learning. Yes, the method itself is not a feature, so further explanation is omitted here.

In the next step S11, the number of repetitions i is incremented, and in the following step S13, the CPU 12 determines whether or not the error (with the same frame prediction as the output) has converged. Then, when "YES" is determined in step S13, this learning process ends.

However, when "NO" is determined in step S13, in step S15, it is determined whether or not the number of repetitions i has reached the specified value. If "NO" in this step S15, the process returns to the previous step S5 and repeats the process of steps S5-S13 described above. If "YES" in step S15, the learning process ends as it is.

The conversion model learned by this learning process is stored in the HDD 20 of FIG. 1 in association with the voice corpus as shown in FIG. However, when it is used for the conversion process of the output voice described later, it is stored in the model storage area 26g (FIG. 3) of the data storage area 26.

FIG. 6 shows each part shown in FIG. 4 as a functional block diagram, and the voice corpus storage unit 101 and the model storage unit 113 correspond to the HDD 20 in the embodiment of FIG. The control rule storage unit 109 corresponds to the control rule storage area 26c in FIG.

Step S1 of FIG. 4 corresponds to the learning parameter generation unit 103, step S5 corresponds to the control parameter generation unit 107, and step S7 corresponds to the learning data generation unit 105. Then, step S9 corresponds to the model learning unit 111.

The operations of these functional units 103, 105, 107 and 111 in FIG. 6 are as described in the corresponding steps S1, S7, S5 and S9, and duplicate description will be omitted here.

In the above embodiment, as the voice corpus, feature amount data based on the voice data of a plurality of (4) utterance styles of each of a plurality of (4) speakers is prepared in advance, and a plurality of utterances in the output voice are prepared. since so as to generate training data in accordance with the number 1 6 with a set of control parameters u _m and v _n is the mixing ratio of styles, by appropriately setting the control parameter u _m and v _n, 7 As shown by the line E of, the degree of emotion (speech style) of the output feature amount with respect to the weight of the input emotion can be designed to change linearly, for example. Therefore, the desired utterance style of the output voice can be easily designed. On the other hand, conventionally, as shown by the line C, the emotional degree (utterance style) of the output feature amount with respect to the weight of the input emotion changes abruptly, so a desired utterance style of the output voice is designed. That was not easy.

However, in the above embodiment, a voice corpus containing voice data of a plurality of utterance styles of a plurality of speakers is prepared, and input is performed using a control parameter including the _{speaker mixing ratio v n} and the utterance style mixing ratio u _m. Data training data (Equation 5) and output data training data (Equation 6) were generated (step S7), and the conversion model was trained in step S9 using the training data. However, the learning data may be generated using a control parameter that includes only the mixing ratio of the utterance styles. In that case, the training data of the input data of the number 5 and the training data of the output data of the number 6 are generated by using only _{the mixing ratio u m of the utterance style, but even in that case, the utterance style in the output voice is generated.} Can be designed.

The output voice conversion model can be learned as described above, and as shown in FIG. 8, the output conversion model can be used as it is, or by introducing an output conversion model constructed by another voice processing device. In addition, it is possible to obtain an output voice obtained by converting the input voice data according to the conversion model.

In the first step S21 of FIG. 8, according to the conversion program 24b (FIG. 3), the CPU 12 captures the input voice waveform and also captures the utterance style parameters u _m and v _n of the output voice desired by the user. The CPU 12 that executes step S21 functions as a waveform and parameter input unit. The captured input waveform is stored in the input voice waveform storage area 26e of the data storage area 26 of the memory 18 shown in FIG. As the input voice waveform, the user's own voice may be input as it is as a waveform signal, or the waveform signal of the voice created by voice synthesis may be input. Further, as an input method, there are a case of inputting from a microphone (not shown) in real time and a method of inputting from a memory such as USB at once.

Further, in order to input the utterance style parameters u _m and v _n , the CPU 12 may display a UI (user interface) (not shown) on the display 23 (FIG. 1) to accept input from the user. and, for example, the parameters u _m and v _n of the desired speech style together with the input voice waveform on the USB described above also may be set, it is also possible to capture from the USB.

In step S23, in a similar manner as step S5 in the previous figures 4, it generates the speech parameters (FIG. 5) from the speech corpus, in a similar manner as step S1, the parameter S _n of a speaker ID Generate. The CPU 12 that executes step S23 functions as a voice parameter generation unit.

In the next step S25, an input parameter such as Equation 5 is generated by combining the voice parameter, the utterance style parameter, and the speaker ID parameter by the same method as in step S7 of FIG. The CPU 12 that executes step S25 functions as an input parameter generation unit. The generated input parameters are stored in the input parameter storage area 26f of the data storage area 26 of the memory 18 shown in FIG.

Then, in step S27, the conversion filter is predicted using the conversion model acquired earlier. The CPU 12 that executes step S27 functions as a conversion filter prediction unit.

In step S29, the conversion filter is used to convert the waveform and output it to the speaker 22 (FIG. 1). The CPU 12 that executes step S29 functions as a waveform conversion and output unit. However, as a waveform conversion method, a method of convolving the input waveform or a method of decomposing the input waveform into the above-mentioned audio parameters, applying the conversion filter, and reorganizing the waveform through the vocoder is applied. ..

According to this embodiment, since at step S21 the user is to set the parameters v _n and u _m of the desired speech style, it is possible to output the output sound from the speech style desired by the user.

FIG. 9 shows each step shown in FIG. 8 as a functional block diagram, and the model storage unit 26g is included in the data storage area 26 of FIG. Step S21 in FIG. 8 corresponds to the waveform and parameter input unit 201, step S23 corresponds to the voice parameter generation unit 203, and step S25 corresponds to the input parameter generation unit 205. Then, Tep S27 corresponds to the conversion filter prediction unit 209, and step S29 corresponds to the waveform conversion unit 211.

However, the operations of these functional units 201, 203, 205, 209 and 111 in FIG. 9 are as already described in the corresponding steps S21, S23, S25, S27 and S29, and duplicate description is omitted here. To do.

10 ... Voice processing device 12 ... CPU
18 ... Memory 20 ... HDD
22 ... Speaker 24 ... Program storage area 26 ... Data storage area

Claims (5)

A voice corpus containing feature data based on the voice data of a calm speech style speech and at least one non-quiet speech style speech,
A control parameter generator that sets a control parameter including a first mixing ratio, which is a mixing ratio of the calm utterance style voice and at least one non-quiet utterance style voice.
From the calm speech style speech feature data and the feature amount of the first input data including control parameters, including the mixing ratio learning data and the undisturbed speech style speech and said at least one undisturbed other speech style audio data A training data generator that generates training data for output data including a conversion filter calculated by a control parameter that includes the first mixing ratio , and input based on the training data for the input data and the training data for the output data. A voice processing device including a learning unit for learning an output conversion model. The speech corpus includes feature data based on the audio data of the undisturbed speech style voice and the at least one undisturbed except for the speech style speech for each of a plurality of different speakers,
The control parameter includes a second mixing ratio, which is a mixing ratio of the plurality of speakers to the same speech style voice.
The learning data generating unit generates the learning data and respective first mixing ratio and second mixing ratio of the speaker and the speech style in accordance including control data, the sound processing apparatus according to claim 1. A conversion filter prediction unit that predicts a conversion filter using the input-output conversion model learned by the speech processing device of claim 1 or 2, and a waveform conversion unit that converts an input waveform using the conversion filter are provided. Voice processing device. A voice processing program that is executed by a computer and uses a voice corpus that includes feature data based on the voice data of a calm speech style speech and at least one non-quiet speech style speech, the processor of the computer.
A control parameter generator that sets a control parameter including a first mixing ratio, which is a mixing ratio of the calm utterance style voice and at least one non-quiet utterance style voice.
From the calm speech style speech feature data and the feature amount of the first input data including control parameters, including the mixing ratio learning data and the undisturbed speech style speech and said at least one undisturbed other speech style audio data A training data generator that generates training data for output data including a conversion filter calculated by a control parameter that includes the first mixing ratio , and input based on the training data for the input data and the training data for the output data. A voice processing program that functions as a learning unit for learning an output conversion model. A speech processing method that is performed by a computer and uses a speech corpus that contains feature data based on speech data of calm speech style speech and at least one non-quiet speech style speech.
A control parameter generation step of setting a control parameter including a first mixing ratio, which is a mixing ratio of the calm utterance style voice and the at least one non-quiet utterance style voice.
From the calm speech style speech feature data and the feature amount of the first input data including control parameters, including the mixing ratio learning data and the undisturbed speech style speech and said at least one undisturbed other speech style audio data A training data generation step that generates training data for output including a conversion filter calculated by a control parameter that includes the first mixing ratio , and input-output based on the training data for the input data and the training data for the output data. A voice processing method that includes learning steps to train a transformation model.

Images