JP6622505B2 - Acoustic model learning device, speech synthesis device, acoustic model learning method, speech synthesis method, program - Google Patents

Description

  The present invention relates to an acoustic model learning device that learns a deep neural network acoustic model from speech data, a speech synthesizer that generates synthesized speech using the learned deep neural network acoustic model, an acoustic model learning method, a speech synthesis method, and a program About.

  A technique based on DNN (Deep Neural Network) is a technique for generating synthesized speech of a speaker from speech data of a target speaker (Non-Patent Document 1). Hereinafter, the configurations and operations of the acoustic model learning device and the speech synthesis device of Non-Patent Document 1 will be described with reference to FIGS. 1 and 2. FIG. 1 is a block diagram showing a configuration of an acoustic model learning device 91 of the same document. FIG. 2 is a block diagram showing the configuration of the speech synthesizer 92 of the same document.

  As illustrated in FIG. 1, the acoustic model learning device 91 of Non-Patent Document 1 includes a speaker voice database 911, an acoustic model learning unit 913, and an acoustic model storage unit 914. The speaker voice database 911 includes a voice data storage unit 9111 and a context data storage unit 9112. The voice data storage unit 9111 stores voice data (voice parameters) of the target speaker in advance. The context data storage unit 9112 stores in advance context data corresponding to the voice data of the target speaker. Although details will be described later, it is assumed that the context data includes at least phoneme information and accent information of the voice data.

  The acoustic model learning unit 913 learns the acoustic model of the target speaker using DNN (Deep Neural Network) using the speech data and context data of the target speaker, and learns the acoustic model (hereinafter referred to as DNN acoustic model, Or the deep neural network acoustic model) is stored in the acoustic model storage unit 914.

  As illustrated in FIG. 2, the speech synthesizer 92 of Non-Patent Document 1 includes a text analysis unit 921, a speech parameter generation unit 922, and a speech waveform generation unit 923.

  The text analysis unit 921 analyzes the input text (text data intended for speech synthesis) and acquires the above-described context data. The voice parameter generation unit 922 generates a voice parameter from the context data using the deep neural network acoustic model stored in the acoustic model storage unit 914. The voice waveform generation unit 923 generates a voice waveform using the generated voice parameter.

Zen et al., "Statistical parametric speech synthesis using deep neural networks." Acoustics, Speech and Signal Processing (ICASSP), 2013 IEEE International Conference on.IEEE, 2013 pp. 7962-7966.

  In order to achieve high quality speech synthesis using the DNN acoustic model, the acoustic model learning unit 913 needs a large amount of speech data and context data of the target speaker. In addition, only a single speaker's voice can be synthesized from one DNN acoustic model.

  For this reason, in order to achieve speech of a plurality of speakers by speech synthesis based on DNN, a large amount of speech data and context data are required for each of the plurality of speakers, and learning costs are high.

  In addition, in order to obtain synthesized speech of a plurality of speakers, it is necessary to store a number of DNN acoustic models corresponding to the number of the speakers, and the number of used memories increases as the number of speakers increases.

  Therefore, an object of the present invention is to provide an acoustic model learning apparatus capable of learning a DNN acoustic model that can generate synthesized speech of a plurality of speakers with a small size at a low cost.

  The acoustic model learning device according to the present invention includes voice data of a plurality of speakers, context data of a plurality of speakers including at least phoneme information and accent information of the voice data, data for specifying a speaker, or speaker characteristics. A deep neural network acoustic model for generating speech parameters necessary for speech waveform synthesis is learned using the data to be represented.

  According to the acoustic model learning device of the present invention, it is possible to learn a DNN acoustic model that can generate synthesized speech of a plurality of speakers with a small size at a low cost.

The block diagram which shows the structure of the acoustic model learning apparatus of a nonpatent literature 1. FIG. The block diagram which shows the structure of the speech synthesizer of a nonpatent literature 1. FIG. 1 is a block diagram illustrating a configuration of an acoustic model learning device according to Embodiment 1. FIG. 3 is a flowchart illustrating the operation of the acoustic model learning device according to the first embodiment. 1 is a block diagram illustrating a configuration of a speech synthesizer according to a first embodiment. 3 is a flowchart showing the operation of the speech synthesis apparatus according to the first embodiment. FIG. 5 is a block diagram illustrating a configuration of an acoustic model learning device according to a second embodiment. 9 is a flowchart illustrating the operation of the acoustic model learning device according to the second embodiment. FIG. 3 is a block diagram illustrating a configuration of a speech synthesizer according to a second embodiment. 9 is a flowchart showing the operation of the speech synthesizer according to the second embodiment. FIG. 9 is a block diagram illustrating a configuration of an acoustic model learning device according to a third embodiment. 10 is a flowchart illustrating the operation of the acoustic model learning device according to the third embodiment. FIG. 9 is a block diagram illustrating a configuration of a speech synthesizer according to a third embodiment. 10 is a flowchart showing the operation of the speech synthesizer according to the third embodiment.

  Hereinafter, embodiments of the present invention will be described in detail. In addition, the same number is attached | subjected to the structure part which has the same function, and duplication description is abbreviate | omitted.

  Hereinafter, the configuration and operation of the acoustic model learning apparatus according to the first embodiment will be described with reference to FIGS. 3 and 4. FIG. 3 is a block diagram illustrating a configuration of the acoustic model learning device 11 according to the present embodiment. FIG. 4 is a flowchart showing the operation of the acoustic model learning device 11 of this embodiment. The difference from the acoustic model learning device 91 of Non-Patent Document 1 is that the acoustic model learning device 11 of the present embodiment uses data for identifying a speaker.

  As shown in FIG. 3, the acoustic model learning device 11 of this embodiment includes a multi-speaker speech database 111, an acoustic model learning unit 113, and an acoustic model storage unit 914. The multi-speaker voice database 111 stores voice data storage units 1111-1,... That store voice data of each speaker for each of a plurality of speakers (N is an integer of 2 or more, N speakers). 1111-N and a context data storage unit 1112-1, ..., 1122-N that stores context data corresponding to each speaker's voice data. The voice data is data of voice in which a plurality of sentences are spoken by N speakers targeted for learning a DNN acoustic model for voice synthesis. The context data is information such as pronunciation given to each utterance in the voice data. The context data holds speech information of speech data, and includes at least phoneme information (pronunciation information) and accent information (accent type, accent phrase length). In addition to this, the context data may include part-of-speech information. The acoustic model storage unit 914 is the same as the component of the same name in the acoustic model learning device 91 of Non-Patent Document 1 described above.

  The acoustic model learning unit 113 generates a DNN acoustic model for generating speech parameters necessary for speech waveform synthesis using speech data of a plurality of speakers and corresponding context data, and data specifying the speakers. Learning and learning the learned DNN acoustic model in the acoustic model storage unit 914 (S113). The data for specifying a speaker is information (data) for specifying a speaker who has read a certain voice data. For example, a speaker code expressing data specifying a speaker as a numerical vector can be used. The speaker code can be a vector in which information for identifying which speaker among the N speakers is expressed in 1-of-K expression. The 1-of-K expression is an expression in which only one element of a vector is 1 and all other elements are 0.

  That is, the acoustic model learning unit 113 receives as input a language feature amount vector representing context data as a numerical vector and a speaker code, and outputs a speech parameter corresponding to the speaker and the context data. Is learned (S113).

  Hereinafter, the configuration and operation of the speech synthesizer 12 of this embodiment will be described with reference to FIGS. FIG. 5 is a block diagram showing the configuration of the speech synthesizer 12 of this embodiment. FIG. 6 is a flowchart showing the operation of the speech synthesizer 12 of this embodiment. The difference from the speech synthesizer 92 of Non-Patent Document 1 is that the speech synthesizer 12 of the present embodiment uses data for specifying a speaker.

As shown in FIG. 5, the speech synthesizer 12 of this embodiment includes a text analysis unit 921, a speech parameter generation unit 122, and a speech waveform generation unit 923. The text analysis unit 921 and the speech waveform generation unit 923 perform the same operations as the components of the same name in the speech synthesizer 92 of Non-Patent Document 1 described above. The voice parameter generation unit 122 uses the DNN acoustic model stored in the acoustic model storage unit 914 to analyze the input text, and the data (speaker) specifying the speaker input together with the input text. Voice parameters are generated from the user code) (S122). The voice parameters include a pitch parameter (basic frequency F0, etc.) and a spectrum parameter (cepstrum, mel cepstrum, etc.). Specifically, the speech parameter generation unit 122 connects the context data and the speaker code, and obtains an input vector to the DNN acoustic model. The voice parameter generation unit 122 inputs the input vector to the DNN acoustic model, and generates a voice parameter by forward propagation (S122). The speech waveform generation unit 923 obtains synthesized speech by speech waveform generation from speech parameters as in Non-Patent Document 1 (S923). The speech waveform generation unit 923 may obtain a speech parameter series smoothed in the time direction using, for example, a maximum likelihood generation (MLPG) algorithm (reference non-patent document 1) before speech waveform generation. For example, (Reference Non-Patent Document 2) may be used for the speech waveform generation.
(Reference Non-Patent Document 1: Mashiko et al., “HMM-based speech synthesis using dynamic features”, Theory of Science, vol.J79-D-II, no.12, pp.2184-2190, Dec. 1996. )
(Reference Non-Patent Document 2: Imai et al., “Mel Logarithmic Spectrum Approximation (MLSA) Filter for Speech Synthesis”, IEICE Transactions A Vol.J66-A No.2 pp.122-129, Feb. 1983 .)

  According to the acoustic model learning device 11 of the present embodiment, in addition to the context data, data (speaker code) for specifying a speaker is utilized, so that the corresponding context data and speech parameters reflecting the speaker characteristics are output. DNN An acoustic model can be learned.

  In the present embodiment, it is assumed that a component that characterizes a speaker and a component that is common among speakers as Japanese speech are included in the speech parameters. Specifically, a speaker code that is a 1-of-K representation of each speaker is used as an input corresponding to a component that characterizes the speaker, and an input corresponding to a component common among the speakers as Japanese speech Context data is used as By providing a speech parameter composed of a component characterizing the speaker and a component common to the speakers as a teacher signal, a parameter estimator corresponding to each component is learned inside the DNN. This enables speech synthesis corresponding to each speaker used for learning with a single DNN acoustic model.

  Since Japanese speech has various speech parameter expressions for various contexts, it was normal to require a large amount of speech data to accurately estimate speech parameters for various contexts. . However, in this embodiment, since it is assumed that the speech parameter includes a component that characterizes the speaker and a component that is common among speakers as Japanese speech, a sufficient amount of speech data exists across multiple speakers. It is only necessary to prepare a large amount of voice data for a single speaker. That is, since voice data of a plurality of speakers is efficiently used and one DNN acoustic model is learned, the voice data necessary for learning can be reduced. Further, since speech synthesis reflecting a plurality of speaker characteristics is realized with one acoustic model, it is possible to realize a speech synthesis system that handles a large number of speakers with less memory usage.

  When the speaker code (1-of-K expression) is used as in the first embodiment, speech synthesis of speakers other than the speakers included in the multi-speaker speech database 111 cannot be performed. Therefore, in the second embodiment, the feature of spectral information of the reference utterance of the target speaker is extracted and used for model learning / speech synthesis, thereby enabling speech synthesis for any target speaker from which the reference utterance is obtained. . Hereinafter, the configuration and operation of the acoustic model learning device 21 according to the second embodiment will be described with reference to FIGS. 7 and 8. FIG. 7 is a block diagram illustrating a configuration of the acoustic model learning device 21 according to the present embodiment. FIG. 8 is a flowchart showing the operation of the acoustic model learning device 21 of the present embodiment. The difference from the acoustic model learning device 11 of the first embodiment is that the acoustic model learning device 21 of the present embodiment uses data (speaker spectrum feature vector) representing speaker characteristics.

  As shown in FIG. 7, the acoustic model learning device 21 according to the present embodiment includes a multi-speaker speech database 111, a spectrum feature extraction unit 212, an acoustic model learning unit 213, and an acoustic model storage unit 914. The person voice database 111 and the acoustic model storage unit 914 are the same as the configuration requirements of the same name in the first embodiment.

The spectrum feature extraction unit 212 extracts each speaker's reference utterance from each speaker's voice data storage units 1111-1 to 1111-N, and calculates each speaker's speaker spectrum feature vector from each speaker's reference utterance. Generate (S212). Here, the reference utterance is an utterance by a speaker used at the time of learning or a target speaker at the time of speech synthesis, and has a feature that it is not necessary to transcribe and may be a short utterance. The speaker spectrum feature vector is a numerical vector representing features of spectrum information found in the speech uttered by the speaker. For example, i-vector may be used to generate the speaker spectrum feature vector. For the spectrum feature extraction unit 212, for example, an i-vector extractor may be used by utilizing the knowledge of Reference Non-Patent Document 3.
(Reference Non-Patent Document 3: Dehak, Najim, et al. "Front-end factor analysis for speaker verification." Audio, Speech, and Language Processing, IEEE Transactions on 19.4 (2011): 788-798.)

  Next, the acoustic model learning unit 213 uses the DNN acoustic model using speech data of a plurality of speakers, context data of the plurality of speakers, and a speaker spectrum feature vector that is data representing speaker characteristics. And the learned DNN acoustic model is stored in the acoustic model storage unit 914 (S213).

  Hereinafter, the configuration and operation of the speech synthesizer 22 of the present embodiment will be described with reference to FIGS. 9 and 10. FIG. 9 is a block diagram showing the configuration of the speech synthesizer 22 of this embodiment. FIG. 10 is a flowchart showing the operation of the speech synthesizer 22 of this embodiment. The difference from the speech synthesizer 12 of the first embodiment is that the speech synthesizer 22 of the present embodiment uses data (speaker spectrum feature vector) representing the features of the speaker.

  As shown in FIG. 9, the speech synthesizer 22 of this embodiment includes a text analysis unit 921, a spectrum feature extraction unit 221, a speech parameter generation unit 222, and a speech waveform generation unit 923. The text analysis unit 921 and the speech waveform generation unit 923 are the same as those in the first embodiment. The spectrum feature extraction unit 221 extracts the above-described speaker spectrum feature vector from the reference utterance input together with the text for speech synthesis (S221). As described above, the reference utterance is an utterance by the target speaker.

  The speech parameter generation unit 222 uses the DNN acoustic model stored in the acoustic model storage unit 914 to analyze the input text and the speech parameter from the speaker spectrum feature vector extracted from the reference utterance. Is generated (S222).

  Since the speech synthesizer 12 of the first embodiment uses speaker codes, target speakers not included in the multi-speaker speech database 111 used during acoustic model learning are unknown during acoustic model learning. Can't synthesize speech. In order to solve this problem, in the present embodiment, a vector (speaker spectrum) expressing features of spectrum information of speech uttered by the speaker, such as i-vector used in the fields of speech recognition and speaker identification. Feature vector). As a result, even if the target speaker is not included in the multi-speaker speech database 111, the speech of a speaker that is acoustically similar to the speech of the target speaker is modeled in the acoustic model. If the speaker's reference utterance can be acquired, it is possible to synthesize speech having spectral characteristics close to the target speaker. Therefore, even a target speaker that is not included in the multi-speaker speech database 111 can generate the synthesized speech. As described above, for example, i-vector can be used to generate the speaker spectrum feature vector, but the implementation method of step S212 is not limited to this.

  In the method of the second embodiment, i-vector, which is a representative method for extracting speaker information vectors from utterances, has been proposed for the purpose of speaker adaptation of models in the speaker identification field and the speech recognition field. It has been. In these fields, it is important that the personality of the spectrum information among the personalities appearing in the speech is expressed by a vector. On the other hand, in the speech synthesis field, when extracting the speaker information vector in order to realize the target speaker's speech synthesis, the personality of the prosodic information as well as the individuality of the spectrum information among the individualities that appear in the speech It is important that this is expressed, and this point is considered to be different from the speech recognition problem. Therefore, in the acoustic model learning device 31 of the third embodiment, the data representing the characteristics of the speaker includes the information of F0. Hereinafter, the configuration and operation of the acoustic model learning device 31 according to the third embodiment will be described with reference to FIGS. 11 and 12. FIG. 11 is a block diagram illustrating a configuration of the acoustic model learning device 31 according to the present embodiment. FIG. 12 is a flowchart showing the operation of the acoustic model learning device 31 of the present embodiment. The difference from the acoustic model learning device 21 of the second embodiment is that the acoustic model learning device 31 of the present embodiment uses not only the speaker spectrum feature vector but also the speaker prosody feature vector as data representing the speaker characteristics. It is.

  As shown in FIG. 11, the acoustic model learning device 31 of the present embodiment includes a multi-speaker speech database 111, a spectrum feature extraction unit 212, a prosody feature extraction unit 312, an acoustic model learning unit 313, and an acoustic model storage. The multi-speaker speech database 111, the spectrum feature extraction unit 212, and the acoustic model storage unit 914 are the same as the constituent elements of the same name in the second embodiment.

  The prosodic feature extraction unit 312 extracts each speaker's reference utterance from each speaker's speech data storage units 1111-1 to 1111-N, and obtains the speaker's prosodic feature vector from each speaker's reference utterance. Generate (S312). The speaker prosody feature vector is a vector expressing the personality of prosodic information among the personalities appearing in speech. More specifically, the speaker prosody feature vector is a numerical vector representing the features of prosodic information among the acoustic features found in the speech uttered by the speaker.

  The prosodic feature extraction unit 312 may calculate, for example, the average and variance of the F0 sequence analyzed from the reference utterance, and extract the F0 feature information as a speaker prosodic feature vector. The prosodic feature extraction unit 312 may perform more detailed prosodic feature modeling using the technique of Reference Non-Patent Document 4. (Reference Non-Patent Document 4: Dehak, Najim, Pierre Dumouchel, and Patrick Kenny. "Modeling prosodic features with joint factor analysis for speaker verification." Audio, Speech, and Language Processing, IEEE Transactions on 15.7 (2007): 2095-2103 .)

  Next, the acoustic model learning unit 313 learns the DNN acoustic model using the speech data of the plurality of speakers, the context data of the plurality of speakers, the speaker spectrum feature vector, and the speaker prosody feature vector. Then, the learned DNN acoustic model is stored in the acoustic model storage unit 914 (S313).

  Hereinafter, the configuration and operation of the speech synthesizer 32 of this embodiment will be described with reference to FIGS. FIG. 13 is a block diagram showing the configuration of the speech synthesizer 32 of this embodiment. FIG. 14 is a flowchart showing the operation of the speech synthesizer 32 of this embodiment. The difference from the speech synthesizer 22 of the second embodiment is that the speech synthesizer 32 of the present embodiment uses not only the speaker spectrum feature vector but also the speaker prosody feature vector as data representing the speaker characteristics. .

  As shown in FIG. 13, the speech synthesizer 32 of this embodiment includes a text analysis unit 921, a spectrum feature extraction unit 221, a prosody feature extraction unit 321, a speech parameter generation unit 322, and a speech waveform generation unit 923. Including. The text analysis unit 921, the spectrum feature extraction unit 221 and the speech waveform generation unit 923 are the same as those in the first embodiment. The prosodic feature extraction unit 321 extracts the above-mentioned speaker prosodic feature vector from the reference utterance input together with the text for speech synthesis (S321). As described above, the reference utterance is an utterance by the target speaker.

  The speech parameter generation unit 322 uses the DNN acoustic model stored in the acoustic model storage unit 914 to generate speech from the context data obtained by analyzing the input text, the speaker spectrum feature vector, and the speaker prosody feature vector. A parameter is generated (S322).

  The acoustic features of a speaker can be classified into spectral features and prosodic features. When the speaker spectral feature vector is used as in the second embodiment, the spectral feature among the speaker features is reflected in the synthesized speech, and the prosody feature of the target speaker is not reflected. In this embodiment, it is possible to synthesize speech that also reflects the prosodic features of speakers not included in the multi-speaker speech database 111 by using a vector that also expresses the prosody information of the target speaker. It becomes.

  Note that the acoustic model learning device and the speech synthesizer described in the above embodiments may be used as an acoustic model learning unit and a speech synthesizer, respectively, and the present invention may be realized as a single piece of hardware including these components.

  In addition, since the speaker code, speaker spectrum feature vector, speaker prosody feature vector, etc. described in the above embodiment have a common term that is a vector expressing speaker information, these are the speaker information vector. May be collectively referred to.

<Supplementary note>
The apparatus of the present invention includes, for example, a single hardware entity as an input unit to which a keyboard or the like can be connected, an output unit to which a liquid crystal display or the like can be connected, and a communication device (for example, a communication cable) capable of communicating outside the hardware entity. Can be connected to a communication unit, a CPU (Central Processing Unit, may include a cache memory or a register), a RAM or ROM that is a memory, an external storage device that is a hard disk, and an input unit, an output unit, or a communication unit thereof , A CPU, a RAM, a ROM, and a bus connected so that data can be exchanged between the external storage devices. If necessary, the hardware entity may be provided with a device (drive) that can read and write a recording medium such as a CD-ROM. A physical entity having such hardware resources includes a general-purpose computer.

  The external storage device of the hardware entity stores a program necessary for realizing the above functions and data necessary for processing the program (not limited to the external storage device, for example, reading a program) It may be stored in a ROM that is a dedicated storage device). Data obtained by the processing of these programs is appropriately stored in a RAM or an external storage device.

  In the hardware entity, each program stored in an external storage device (or ROM or the like) and data necessary for processing each program are read into a memory as necessary, and are interpreted and executed by a CPU as appropriate. . As a result, the CPU realizes a predetermined function (respective component requirements expressed as the above-described unit, unit, etc.).

  The present invention is not limited to the above-described embodiment, and can be appropriately changed without departing from the spirit of the present invention. In addition, the processing described in the above embodiment may be executed not only in time series according to the order of description but also in parallel or individually as required by the processing capability of the apparatus that executes the processing. .

  As described above, when the processing functions in the hardware entity (the apparatus of the present invention) described in the above embodiments are realized by a computer, the processing contents of the functions that the hardware entity should have are described by a program. Then, by executing this program on a computer, the processing functions in the hardware entity are realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto-Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, a hardware entity is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

Claims (6)

And audio data of a plurality of speakers, a plurality of speakers context data including phonemic information and accent information of at least the audio data, and data for identifying the speaker represented by 1-of-K expression vector An acoustic model learning unit that learns a deep neural network acoustic model for generating speech parameters including a pitch parameter and a spectrum parameter necessary for speech waveform synthesis using
To the input layer of the neural network, enter at least a plurality of speakers context data including the phoneme information and accent information of the audio data, the data for specifying the speaker represented by 1-of-K expression vector An acoustic model learning device characterized by that. A text analysis unit that analyzes input text and obtains context data including at least phoneme information and accent information;
And audio data of a plurality of speakers, and context data of the speech data of the plurality of speakers, inputs the data for identifying the speaker is represented in the input layer of the neural network 1-of-K expression vector pitch parameter and, and, by using the learning has been deep neural network acoustic models to generate speech parameters including a spectrum parameter, and context data obtained by analyzing the input text, together with the input text data do we pitch parameter that identifies the speaker represented by 1-of-K expression vector is input, and a speech parameter generation unit for generating a speech parameter containing spectral parameter,
A voice waveform generation unit that generates a voice waveform using the generated voice parameter;
To the input layer of the neural network, inputting and context data obtained by analyzing the input text, the data for identifying the speaker represented by 1-of-K expression vector is input together with the input text A speech synthesizer characterized by the above. The speech synthesizer according to claim 2,
The deep neural network acoustic model is
Learned using the voice data, the context data, and data representing the characteristics of the speaker;
The data representing the speaker characteristics include:
A speech synthesizer including at least a speaker prosody feature vector which is a vector expressing features of prosody information of speech uttered by the speaker. An acoustic model learning method executed by the acoustic model learning device,
And audio data of a plurality of speakers, a plurality of speakers context data including phonemic information and accent information of at least the audio data, and data for identifying the speaker represented by 1-of-K expression vector Learning a deep neural network acoustic model for generating speech parameters including pitch parameters and spectral parameters necessary for speech waveform synthesis using
To the input layer of the neural network, enter at least a plurality of speakers context data including the phoneme information and accent information of the audio data, the data for specifying the speaker represented by 1-of-K expression vector An acoustic model learning method characterized by the above. A speech synthesis method executed by a speech synthesizer,
Analyzing the input text to obtain context data including at least phoneme information and accent information;
And audio data of a plurality of speakers, and context data of the speech data of the plurality of speakers, inputs the data for identifying the speaker is represented in the input layer of the neural network 1-of-K expression vector pitch parameter and, and, by using the learning has been deep neural network acoustic models to generate speech parameters including a spectrum parameter, and context data obtained by analyzing the input text, together with the input text 1-of-K data or we pitch parameter that identifies the speaker represented by expression vectors that are input, and a step of generating a speech parameters including spectral parameter,
Generating a speech waveform using the generated speech parameters;
To the input layer of the neural network, inputting and context data obtained by analyzing the input text, the data for identifying the speaker represented by 1-of-K expression vector is input together with the input text A speech synthesis method characterized by the above. The program which makes a computer function as a speech synthesizer of Claim 2 or 3 .

Images