JP2008191477A - Hybrid type speech synthesis method, its device, its program and its recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a hybrid type speech synthesis device by which high quality synthesized speech is stably obtained, in a hybrid type speech synthesis system in which synthesized speech is created by switching over a connecting waveform type speech synthesis method and a Hidden Markov Model (HMM) speech synthesis method. <P>SOLUTION: A connecting waveform type synthesized speech creating section creates synthesized speech data in which a content coincides with an input text by searching for and connecting an elementary speech unit of arbitrary length included in a speech corpus. An HMM synthesized speech creating section creates a statistical model beforehand, by learning speech information extracted from the speech corpus by a statistical method, and calculates a speech information parameter corresponding to the input text from the statistical model, and creates the synthesized voice data from the parameter. The above two speech synthesis systems are included, and a speech quality comparison determination means performs comparison and determines which is higher in speech quality, the synthesized speech data created by the connecting waveform type synthesized speech creating section, or that created by the HMM synthesized speech creating section, for each syllable. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention comprises a waveform-connected speech synthesis method and an HMM speech synthesis method, a hybrid speech synthesis method that outputs high-quality synthesized speech that matches an arbitrary text input, its apparatus, its program, and its program. The present invention relates to a storage medium for storage.

2. Description of the Related Art In recent years, with the advance of speech synthesis technology, speech synthesis is used for reading out information guidance, Web articles, e-mails and the like in an automatic voice response device (IVR). As main speech synthesis methods, a waveform-connected speech synthesis method as disclosed in Patent Document 1 and a hidden Markov model speech synthesis method (Hidden Markov Model, hereinafter referred to as HMM speech synthesis method) disclosed in Non-Patent Document 1. Is mentioned.
The waveform-connected speech synthesis method can generate high-quality synthesized speech when a large-scale speech corpus (Corpus) can be recorded.

On the other hand, the HMM speech synthesis method can generate synthesized speech having a certain level of quality even when a small speech corpus is used.
When it is desired to generate synthesized speech of various tone, it is difficult to create a large-scale speech corpus for each tone, so synthesized speech by the waveform connection type speech synthesis method cannot be generated with high quality. This is because high-quality synthesized speech can be generated where the corresponding segment exists in the speech corpus, but the quality is greatly improved if the desired segment does not exist in the speech corpus. This is because it will deteriorate. However, at present, sufficient speech quality cannot be realized only by the HMM speech synthesis method.

Therefore, by considering a hybrid speech synthesis method that generates synthesized speech by switching whether to use synthesized speech by waveform connection speech synthesis method or synthesized speech by HMM speech synthesis method in a certain basic unit, higher quality is achieved. Non-Patent Document 2 discloses a method for generating a simple synthesized speech from a small-scale speech corpus. In this document, which type of synthesized speech is used is determined based on whether the speech unit to be synthesized is included in a small-scale corpus or not. If the speech unit is used and contains less than a certain number of speech units, the synthesized speech may be deteriorated. Therefore, it is determined based on the selection criterion that the synthesized speech by the HMM speech synthesis method is used.
Further, in Non-Patent Document 2, the determination of which type of synthesized speech is used is performed in units of diphones (phoneme center to next phoneme center) with the phoneme center as a boundary.
Japanese Patent No. 2761552 Keiichi Tokuda, “Basics of Speech Synthesis with HMM”, Shingaku Techniques, SP2000-74, 43-50, Oct. 2000 Masafumi Okubo, Ryo Mochizuki, Tetsunori Kobayashi, “Study of speaker conversion method using HMM segment selection”, Shingaku technique, SP2004-139, pp. 13-18, Jan. 2005

However, it cannot be said that the number of speech units to be synthesized in the small-scale speech corpus is directly related to the speech quality of waveform connected speech synthesis. Since it is possible to consider high quality even if there are few speech units and low quality even if there are many speech units, either speech synthesized by the waveform connection speech synthesis method or synthesized speech by the HMM speech synthesis method It is desirable to determine whether to use a segment after judging the quality of each voice.
The present invention has been made in view of such problems, and it is determined whether each speech unit to be used is synthesized speech by the waveform-connected speech synthesis method or synthesized speech by the HMM speech synthesis method. It is an object of the present invention to provide a hybrid speech synthesis method, a device thereof, a program thereof, and a storage medium that are determined after determining the quality of the speech.

The hybrid speech synthesizer according to the present invention includes a waveform connection type synthesized speech generation unit, an HMM synthesized speech generation unit, speech quality comparison / determination means, and hybrid synthesized speech generation processing means. The waveform connection type synthesized speech generation unit generates synthesized speech data whose contents match the input text by searching for and connecting speech units of arbitrary length included in the speech corpus. The HMM synthesized speech generation unit generates a statistical model in advance by learning speech information extracted from the speech corpus by a statistical method, obtains speech information parameters corresponding to the input text from the statistical model, and synthesizes from the parameters Generate audio data.
The voice quality comparison / determination means determines which voice quality of the synthesized voice data generated by the waveform connection type synthesized voice generator and the synthesized voice data generated by the HMM synthesized voice generator is higher for each syllable unit of the synthesized voice data. The comparison judgment is performed.

The hybrid synthesized speech generation processing means connects the synthesized speech data units according to the determination result of the speech quality comparison / determination means to generate hybrid synthesized speech data.
In addition, the speech corpus of the hybrid speech synthesizer according to the present invention generates speech data designed to improve the quality of synthesized speech data generated by the waveform-connected synthesized speech generation unit and the HMM synthesized speech generation unit. Voice data designed so that the quality of the synthesized voice data is high.

  According to the hybrid speech synthesizer according to the present invention, it is determined whether the quality of the synthesized speech by the waveform connection speech synthesis method or the synthesized speech by the HMM speech synthesis method is higher, and the sense of incongruity due to the difference in sound quality due to the difference in the synthesis method. High-quality synthesized speech data can be used for each syllable unit determined so that there are few discontinuities in the connection points and high-quality synthesized speech even when using a small speech corpus. Can be generated.

  Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated.

  FIG. 1 shows a first embodiment of the hardware configuration of the hybrid speech synthesizer 100 of the present invention. The first embodiment is an example in which a hybrid speech synthesizer 100 is configured using a general personal computer.

  In FIG. 1, reference numeral 102 denotes a control memory (ROM), which stores various control data used by a central processing unit (CPU) 104. The CPU 104 executes a control program stored in the RAM 106 to control the overall operation of the hybrid speech synthesizer 100. The RAM 106 temporarily stores various data used as a work area when the CPU 104 executes various control processes, and loads and stores a control program from the external storage device 108 when the CPU 104 executes various processes. Yes. The external storage device 108 is, for example, a hard disk device or an optical disk drive. The control program is stored in, for example, a DVD (Digital Versatile Disc) or CD-ROM (Compact Disc Read Only Memory) of an optical disc and can be transferred to another personal computer.

  When digital data indicating an audio signal is input, the D / A converter 110 converts the data into an analog signal and outputs the analog signal to the speaker 118 to reproduce the audio. An input unit 112 includes a keyboard 120 and a pointing device such as a mouse (not shown). A display unit 114 includes a display 122 such as a CRT or a liquid crystal. Reference numeral 116 denotes a hybrid speech synthesis unit that generates speech synthesis digital data that matches the input text data. A bus 150 connects the above-described parts.

  In the above configuration, the control program for controlling the hybrid type speech synthesis unit 116 of the embodiment of the first embodiment is loaded from the external storage device 108 and stored in the RAM 106, and various data used in the control program is controlled by the control program. Stored in the memory 102. These data are appropriately taken into the RAM 106 through the bus 150 under the control of the CPU 104 and used in the control process by the CPU 104. The D / A converter 110 converts the speech synthesis digital data created by the hybrid speech synthesis unit 116 by executing a control program into an analog signal and outputs the analog signal to the speaker 118.

  FIG. 2 shows a functional configuration example of the module of the hybrid type speech synthesis unit 116 of the first embodiment, and FIG. 3 shows an operation flow thereof. Text data is input as a data file from the input unit 112 or from the external storage device 108 (step S112). The input text data is morphologically analyzed by the text analysis processing unit 32, decomposed into parts of speech, and given a reading and an accent type (step S32). The text analysis processing unit 32 outputs a reading and an accent type by using a syllable as one synthesized speech data unit. A target prosody pattern based on the reading and accent type is generated by the prosody information generation unit 34 (step S34). Speech synthesis data that matches the target prosodic pattern as much as possible is generated by the synthesis processing unit 40 (step S40).

The operation of assigning reading and accent type in the text analysis processing unit 32 (step S32) and the generation of the target prosody pattern in the prosody information generation unit 34 (step S34) are the same as those in the conventional speech synthesizer.
A main part of the present invention is a synthesis processing unit 40. The synthesis processing unit 40 generates a synthesized voice data unit (syllable) by the waveform connection type voice synthesis method and a synthesized voice data unit by the HMM voice synthesis method (steps S42 and S44), and the voice quality of either synthesized voice data unit. Is compared for each syllable unit (step S50) to generate synthesized speech with little quality degradation.

If the unit for comparing and selecting the quality of the synthesized speech by both methods is too long, the HMM synthesized speech that is not so high in quality is clearly perceived, so the quality of the output synthesized speech in which the synthesized speech of both types is mixed is It will be lower. On the other hand, when the unit to be selected is too short, the number of connection points of the synthesized speech by both types increases, resulting in frequent voice deterioration due to discontinuities at the connection points. Therefore, the syllable unit is used as a unit in which it is difficult to perceive the quality of the HMM synthesized speech itself and the number of connection points of synthesized speech by both systems does not increase too much.
As a result, connecting the higher-quality synthesized speech data in syllable units results in fewer connection points than phonemes or diphone units, so that deterioration due to discontinuities in connection points hardly occurs. Further, by suppressing the continuous output of HMM synthesized speech with insufficient speech quality, it is possible to prevent quality degradation caused by clearly perceiving HMM synthesized speech that is not so high in quality.

The synthesis processing unit 40 includes a speech corpus 42 that is a database created by mixing speech data that improves the synthesized speech quality of the synthesized speech by the waveform connection speech synthesis method and the synthesized speech by the HMM speech synthesis method, The waveform connection type synthesized speech generation unit 44, the signal processing unit 46, the HMM synthesized speech generation unit 48, the speech quality comparison / determination unit 50, and the hybrid synthesis speech generation processing unit 52 are configured.
The waveform connection type synthesized speech generation unit 44 searches the speech corpus 42 to generate synthesized speech data that is a speech unit sequence as close as possible to the target prosody pattern input from the prosody information generation unit 34.

  Since the synthesized speech data by the waveform connected speech synthesis method is obtained by connecting speech units, it is not guaranteed to reproduce the target prosodic pattern. Therefore, the signal processing unit 46 performs signal processing on the prosody of the synthesized speech data of the waveform connection type speech synthesis method to accurately match the target prosody pattern. Signal processing techniques performed by the signal processing unit 46 include a pitch-synchronized waveform superposition method (PSOLA method) in which the prosody is changed by manipulating the pitch in the time domain, and the pitch in the frequency domain. To change the prosody.

  The HMM synthesized speech generation unit 48 includes a statistical model learning unit 48b, a speech parameter conversion unit 48a, and a statistical model 48c. The statistical model learning unit 48b creates a statistical model 48c in advance based on the speech data of the speech corpus 42. The speech parameter conversion unit 48a converts speech parameters obtained by inputting the target prosody pattern input from the prosody information generation unit 34 into the statistical model 48c into synthesized speech data. The operation of generating the synthesized speech data by the waveform connection type synthesized speech generating unit 44 and the HMM synthesized speech generating unit 48 is not different from that of a speech synthesizer conventionally provided with each method alone.

  In contrast to the conventional speech synthesizer, the first embodiment is characterized in that the waveform connection type synthesized speech generation unit 44 and the HMM synthesized speech generation unit 48 share one speech corpus 42, and each speech generation. The new feature is that the quality of the synthesized speech data in syllable units generated by the sections is compared, and the synthesized speech data having the higher quality is output.

[Voice Corpus 42]
The speech corpus 42 is speech data designed and collected so that the quality of synthesized speech by the waveform connection type speech synthesis method is high, and speech that is designed and collected so that the quality of synthesized speech by the HMM speech synthesis method is high. Data is a mixed database.
In order to increase the quality of synthesized speech by the waveform-connected speech synthesis method, it is necessary to use a speech database that preferentially includes speech data that is frequently used for synthesis. In Japanese, speech data frequently used for synthesis means speech data including a common wording or word, frequently used syllables, phonology, phonological environment, fundamental frequency, power, and the like with high probability.

  On the other hand, in order to improve the quality of synthesized speech by the HMM speech synthesis method, since the HMM is a statistical model, the number of data for learning the HMM model must be uniform to some extent for each learning unit. The learning unit of the HMM model is divided according to a spectrum, which is one of speech information, that is, a phoneme and a phoneme environment. Therefore, high-quality HMM synthesized speech can be created by using speech data collected taking into consideration only the balance of phonemes and phonemic environments without considering the fundamental frequency.

A method for constructing a speech corpus storing such speech data is disclosed in, for example, Japanese Patent Application Publication No. 2004-246140. As a specific construction method of the speech corpus 42, for example, about half of predetermined speech data, the acoustic and linguistic weights are made the same, and the text is selected in consideration of the Japanese frequency to some extent. For the other half, for example, a simple text is selected so that the number of triple phonemes (triphones) having a high frequency in Japanese is the same. By constructing the speech corpus 42 in this way, it is possible to collect texts that are balanced with a suitable phoneme and phoneme environment for an HMM speech synthesis method even with a relatively small database. In addition, since it is composed of text collected with emphasis on the frequency of Japanese, it is possible to provide a speech corpus 42 in which prosodic and phoneme variations important for waveform-connected speech synthesis methods are guaranteed.
The respective synthesized speech data generated by the waveform connection type synthesized speech generation unit 44 and the HMM synthesized speech generation unit 48 based on the speech data of the speech corpus 42 are input to the speech quality comparison / determination means 50.

[Audio quality comparison / determination means 50]
The voice quality comparison / determination means 50 determines which voice quality is higher, the synthesized voice data unit generated by the waveform connection type synthesized voice generating unit 44 or the synthesized voice data unit generated by the HMM synthesized voice generating unit 48. Comparison is made for each unit.
As a speech quality comparison method, a method of determining by a statistical method such as machine learning, a method of determining using a threshold value, a method of determining by designing a cost function, and the like can be considered. As an example of the determination method, an example in which a cost function using a speech physical quantity parameter as an input will be described in the first embodiment.

The voice quality comparison / determination unit 50 includes a cost value conversion unit 50a that converts a physical quantity parameter into a cost value, and a comparison / determination unit 50b that compares the cost value. The physical quantity parameter used for determination differs depending on the speech synthesis method.
In the waveform-connected synthetic speech method, since significant deformation of the speech waveform due to signal processing leads to quality degradation, the average value of the difference between the target prosody pattern and the prosody pattern of the speech segment sequence is used as a parameter for determining the degree of quality degradation. The maximum value can be used. Further, as a physical quantity parameter for determining the ambiguity of a speech segment, a value obtained by comparing the prosodic environment in the speech corpus of the selected speech segment with the prosody environment of the input text can be used.

  On the other hand, in the HMM synthesized speech method, for example, the number of phoneme data included in a learning unit when creating a statistical model can be used. This is because when the number of phoneme data included in the learning unit is small, it is difficult to create a statistical model with high reliability, and it is considered that the quality of synthesized speech by the HMM speech synthesis method is lowered.

  In addition, by using as a parameter the proportion of the phonetic environment of the input text and the corresponding phonological environment of the data used for HMM learning in syllable units, it is determined whether the statistical model is highly reliable. it can. If the phonological environments do not match, the sound characteristics change even for the same syllable. It is ideal to learn only with data that has the same phonological environment, but if there is no such data, HMM may be learned using data that ignores the phonological environment. Thus, the quality of the HMM synthesized speech can be determined.

  The number of continuous selections of HMM synthesized speech when HMM synthesized speech is selected for the corresponding syllable can also be used as a parameter. This is because, in this method of selecting between waveform-connected synthesized speech and HMM synthesized speech in syllable units, HMM synthesized speech that is not so high in quality is clearly perceived when HMM synthesized speech is selected continuously. This is because the quality of the output synthesized speech in which both types of synthesized speech are mixed is lowered. That is, it is a parameter for the purpose of controlling so that HMM synthesized speech is not continuously selected.

  Since the dimensions of the physical quantity parameters as described above are different, they cannot be simply compared. Therefore, in Example 1, each physical quantity is converted into a cost value using a cost function (formula (1)) and then compared.

ここで、
Ｓ：サブコスト関数
ｆ：シグモイド関数
ｘ：判定のために入力される物理量
θ：物理量に対する閾値
ｉ：物理量パラメータの種類
ｗ：各物理量パラメータに対する重み
α：シグモイド関数の傾斜調節パラメータ
シグモイド関数は、図４に示すようにｙ＝０とｙ＝１を漸近線に持ち、ｅの指数部が０の時にｙ＝１/２を示す関数である。横軸ｘを物理量パラメータとすることで、次元の異なる物理量パラメータを一つの指標であるコスト値に変換出来る。コスト値に変換する際に、各物理量の指数部の各値、閾値θ、重みｗ、傾斜調節パラメータαを操作することで、各物理量パラメータの比較水準を調節することが出来る。この調節は、実験等の結果から事前に設計しておく。

here,
S: Sub-cost function f: Sigmoid function x: Physical quantity input for determination θ: Threshold for physical quantity i: Kind of physical quantity parameter w: Weight for each physical quantity parameter α: Slope adjustment parameter of sigmoid function The sigmoid function is shown in FIG. As shown in the figure, y = 0 and y = 1 are asymptotic lines, and when the exponent part of e is 0, the function indicates y = 1/2. By using the horizontal axis x as a physical quantity parameter, physical quantity parameters having different dimensions can be converted into a cost value as one index. When converting to a cost value, the comparison level of each physical quantity parameter can be adjusted by manipulating each value of the exponent part of each physical quantity, threshold value θ, weight w, and slope adjustment parameter α. This adjustment is designed in advance based on the results of experiments and the like.

FIG. 5 shows an operation flow of the voice quality comparison / determination process (step S50) performed by the voice quality comparison / determination means 50. The cost value conversion unit 50a in the speech quality comparison / determination unit 50 inputs a target prosodic pattern for each synthesized speech data unit input from the waveform-connected synthesized speech generation unit 44 and a synthesized speech data unit that is a speech unit sequence. As described above, the difference Rd between the target prosodic pattern for each unit and the prosodic pattern formed by the speech segment is converted into the sub-cost value S ₁ using the equation (1) (step S520). Also, in terms of the agreement degree _{A L} of the synthesized speech data units before and after the phoneme environment as a sub-cost value _{S 3} (step S521). Cost 1 is calculated by adding the sub-cost values S ₁ and S ₃ (step S522).

Cost value of the synthesized speech data units output by the HMM speech synthesizer unit 48, the cost value conversion unit 50a, sub-cost value S ₂ that depends on the number N of learning data included in the synthesized speech data units is calculated ( Step S523). Further, it converted as sub-cost value _{S 4} also HMM synthesized speech data to the matching degree _{A HL} synthesized speech data units before and after the phoneme environment (step S524). Adding the sub-cost value _{S 2} and _{S 4} to calculate the Cost2 (step S525).

Cost1 and Cost2 are compared by the comparison determination unit 50b (step S526). In the case of Cost1 ≦ Cost2 (step S526, Yes), the voice quality of the synthesized voice data unit generated by the waveform connection type synthesized voice method is better than the synthesized voice data unit generated by the HMM synthesized voice method. I can judge.
When Cost1> Cost2 (step S526, No), it can be determined that the voice quality in units of synthesized voice data generated by the HMM synthesized voice method is better.

The synthesized speech data unit of the synthesized speech method determined to have good speech quality in the comparison determination unit 50 b is output to the hybrid synthesized speech generation processing means 52.
The hybrid synthesized speech generation processing means 52 connects the synthesized speech data units that are sequentially input, and outputs the hybrid synthesized speech data to the bus 150 (step S529).

[Modification]
Alternatively, one cost value depending on each of the waveform connection type synthesized speech data and the HMM type synthesized speech data is calculated, and the synthesized speech data is selected by comparing the one cost value with a threshold value. Good. FIG. 6 shows an operation flow of a modified example (step S50 ′) of the voice quality comparison determination process.
The cost function used here represents the degree of barrier (cost) to using the HMM synthesized speech, and the waveform connected synthesized speech has a higher cost value when the quality is higher, and the HMM speech synthesized has a higher quality. The case is a cost function designed to output a low cost value.

Here, a target prosody pattern for each unit of synthesized speech data input from the waveform connection type synthesized speech generation unit 44 and a synthesized speech data unit that is a speech unit sequence are input, and the target prosody pattern and speech unit for each unit are input. The difference Rd from the formed prosodic pattern is converted into the sub cost value S ₁ by the expression (1) (step S520), and the degree of agreement A _L of the phoneme environment before and after the synthesized speech data unit is used as the sub cost value S ₃ as the sub cost. converted into the value _{S 3} (step S521) the processing is different from FIG. In step S520 and S521, if the voice quality of the synthesized speech data of concatenative high, calculates a larger sub-cost values S ₁ and S _3, if the voice quality is low reversed to calculate the small sub-costs value.

On the other hand, the cost value of the synthesized speech data units output by the HMM speech synthesizer unit 48, as the number N is large training data contained in the synthesized speech data units, and calculates a smaller sub-cost value S ₂ (step S523 ') . Further, the matching degree A _HL of the phoneme environment before and after the synthesized speech data unit of the HMM synthesized speech method is also expressed by the equation (1) so that the higher the matching degree, the smaller the sub cost value S ₄ is converted (step S524 ′). Design a sigmoid function.

When such sub cost values S _{1 to} S ₄ are calculated as one cost value Cost 1 (step S 522), if the speech quality of the synthesized speech data of the HMM synthesized speech method is high, the cost value Cost 1 shows a small value, The cost value Cost1 (score value) indicates a large value when the speech quality of the synthesized speech data of the waveform connection speech synthesis method is high.
By comparing the cost value Cost1 designed in this way with a threshold value, it is possible to determine which type of synthesized voice data has higher voice quality.

  As described above, according to the present invention, it is determined whether the quality of the synthesized speech by the waveform connection type speech synthesis method or the synthesized speech by the HMM speech synthesis method is high, and synthesis of a high quality method is performed for each unit of speech synthesis. Since voice data can be used, high-quality synthesized voice can be generated even when a relatively small voice database is used.

  In addition, although the example using a cost function was shown and demonstrated as a speech quality comparison method, it is an example and another method is also considered. For example, the comparison determination may be performed using a statistical method such as machine learning. One of the machine learners is determined based on the result of the determination and the above-described physical quantity parameter by determining which of the synthesized speech data unit of the waveform connection type speech synthesis method and the synthesized speech data unit of the HMM speech synthesis method is better. By creating a binary classifier (SVM: Support Vector Machine) and inputting an unknown physical quantity parameter to the binary classifier, it is possible to determine which method to use synthesized speech. May be.

  In addition, the hybrid speech synthesizer of the present invention has been described with an example in which a general personal computer is used to configure software, but each module of the hybrid speech synthesizer 116 shown in FIG. It is also possible to configure with hardware.

The figure which shows Example 1 which is an example of the hardware constitutions of the hybrid type speech synthesizer 100 of this invention. FIG. 3 is a diagram illustrating an example of a functional configuration of a module of the hybrid type speech synthesis unit 116 according to the first embodiment. The figure which shows the operation | movement flow of the hybrid type speech synthesis unit 116. The figure which shows a sigmoid function. The figure which shows the operation | movement flow of the audio | voice quality comparison determination means 50 (step S50). The figure which shows the operation | movement flow of the modification (step S50 ') of the audio | voice quality comparison determination means 50. FIG.

Claims (5)

A waveform-connected synthesized speech generation unit that generates synthesized speech data whose content matches the input text by searching for and connecting speech units of arbitrary length included in the speech corpus;
A statistical model is generated in advance by learning speech information extracted from the speech corpus using a statistical method, speech information parameters corresponding to the input text are obtained from the statistical model, and synthesized speech data is generated from the speech information parameters. A hybrid speech synthesizer comprising an HMM synthesized speech generation unit
Compare and determine for each syllable unit of the synthesized speech data which speech quality is higher, the synthesized speech data generated by the waveform-connected synthesized speech generation unit and the synthesized speech data generated by the HMM synthesized speech generation unit Voice quality comparison / determination means,
A signal processing unit that performs processing for matching a syllable unit prosody of the waveform-connected speech synthesis method to a target prosody pattern between the waveform-connected synthesized speech generation unit and the speech quality comparison and determination unit;
Hybrid synthesized speech generation processing means for generating hybrid synthesized speech data by connecting the synthesized speech data units according to the determination result of the speech quality comparison determining means;
A hybrid type speech synthesizer characterized by comprising: The hybrid speech synthesizer according to claim 1,
The speech corpus is designed so that the quality of the synthesized speech data generated by the waveform connection type synthesized speech generation unit is high and the quality of the synthesized speech data generated by the HMM synthesized speech generation unit is high. A hybrid type speech synthesizer characterized by comprising: voice data designed as described above. A waveform connected synthetic speech generation process for generating synthesized speech data whose content matches the input text by searching for and connecting speech segments of arbitrary length included in the speech corpus,
Obtaining an audio information parameter corresponding to the input text and generating synthesized speech data from the speech information parameter;
A signal processing process for processing the syllable unit prosody of the waveform connected speech synthesis method to match the target prosody pattern,
A speech quality comparison / determination process for comparing each syllable unit to determine which voice quality is higher, the syllable unit generated in the waveform connected synthetic speech process or the syllable unit generated in the HMM synthesized speech generation process; ,
A hybrid synthesized speech generation process for generating hybrid synthesized speech data by connecting the synthesized speech data units according to the determination result of the speech quality comparison determination process;
A hybrid speech synthesis method characterized by comprising:   An apparatus program for causing a computer to function as each apparatus according to claim 1.   A computer-readable storage medium storing any one of the programs according to claim 4.

Images