JP3554513B2 - Speech synthesis apparatus and method, and recording medium storing speech synthesis program - Google Patents

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a speech synthesizer that generates speech from text.
[0002]
[Prior art]
FIG. 1 shows a configuration example of a general speech synthesizer. The speech synthesizer shown in FIG. 1 includes an input terminal (101), a text analyzer (102), a prosody generator (103), a speech unit selector (104), a speech unit database (105), a speech synthesizer ( 106 ) and an audio output terminal (107).
[0003]
The conventional speech synthesizer ( 106 ) is an example of a speech synthesizer using a parameter synthesis method based on a linear prediction method. As shown in FIG. 5 , the voice synthesizer includes a pulse generator (201), a noise generator (202), a sound source switcher (203), an amplifier (204), a synthesis filter (205), an output terminal (206), It comprises an inverse quantizer (501) and a parameter quantization table (502).
[0004]
Hereinafter, referring to FIG. 1, the operation of the conventional speech synthesis apparatus. The text analyzer (102) inputs text information (eg, “left”) such as words and sentences, and outputs reading information (eg, hidari).
[0005]
The prosody generator (103) receives reading information (eg, hidari) and generates prosody information (pitch, loudness, duration of voice). The pitch of the voice is determined by the pitch frequency of the vowel. In the case of an example, the pitch frequency of the vowel i, a, i is determined in chronological order. As for the volume and the utterance speed of the voice, amplitude information and duration time are determined for each phoneme (eg, h, i, d, a, r, i).
[0006]
The speech unit selector (104) refers to the speech unit database (105) and selects a speech unit based on reading information (eg, hidari). As a speech synthesis unit, a consonant + vowel (CV: Consonant, Vowel) phoneme unit (eg, ka, gu) or a vowel + consonant + vowel (VCV) holding a feature amount of a phoneme transition portion for the purpose of improving sound quality. ) Units (eg, aki, ito) are widely used.
[0007]
For example, Japanese Patent Application Laid-Open No. 60-208799 describes a method of storing data in units of phonemes in a synthesizing system using a parameter synthesizing method. In this system, vocal tract information (linear prediction coefficients) of speech is stored in a phonological memory as one set of data (for example, in hi, a linear prediction coefficient of a consonant part h and a vowel part i).
[0008]
In the following description, a method using a VCV unit will be described. The speech unit database (105) analyzes speech data appropriately cut out from natural speech data in units of VCV, and holds information necessary for synthesis. In the case of Japanese, the number of VCV segments is about 1000, and the time length of each segment may be normalized and held or may be different.
[0009]
The audio signal is analyzed for each predetermined length (frame). The information held for each VCV includes phoneme boundary position information of the VCV (seg1, seg2, seg3 = VCV frame number), and voiced / unvoiced time-series information (uv [1], uv [2],...) For each frame. , uv [seg3]), amplitude time-series information for each frame (pow [1], pow [2],..., pow [seg3]), and time-series information (a [1]) for the linear prediction coefficient vector for each frame , a [2],..., a [seg3]). Here, as the linear prediction coefficient vector (a [k], k = 1,..., Seg3), a linear prediction coefficient of about the tenth order, a line spectrum pair parameter, and the like are used.
[0010]
The operation of the speech unit selector (104) will be described in the above example (h, i, d, a, r, i). As the VCV unit to be selected, * hi, ida, ari, i ** is selected. And output as speech unit information. Here, the symbol * represents silence. In the synthesis method using VCV as a unit, adjacent segments are connected by vowels in this way.
[0011]
Finally, the speech synthesizer ( 106 ) inputs prosody information and speech unit information, and generates and outputs synthesized speech while controlling voice pitch, loudness, and utterance speed according to the prosody information. I do.
[0012]
In FIG. 5, a sound source switch (203) switches sound sources based on voiced / unvoiced information of speech unit information so as to select a pulsed sound source when voiced and a noise source when voiceless. For voiced, pulse generator (201), based on the voice of prosodic information height (pitch frequency), generates a pulse train at appropriate time intervals. In the case of unvoiced sound, a white noise is generated by a noise generator (202). The amplifier (204) appropriately amplifies the amplitude of the sound source signal based on the loudness (amplitude information) of the voice of the prosody information. In general, a speech synthesizer using a parameter synthesis method appropriately quantizes a linear prediction coefficient and then stores it in a speech unit database (105). The speech synthesizer appropriately decodes the linear prediction coefficient parameter in the speech unit information by the inverse quantizer (501) using the parameter quantization table (502), and then sends the decoded result to the synthesis filter (205). send. The synthesis filter (205) sets the decoded linear prediction coefficient as a filter coefficient, performs a filtering process on the amplified sound source signal, and outputs the result as an audio waveform.
[0013]
In the above description, the parameter synthesizer-based speech synthesizer has been described, but a waveform editing system (for example, Japanese Patent Laid-Open No. 60-21098) may be used as the speech synthesizer. Hereinafter, a speech synthesizer using the waveform editing method will be described.
[0014]
The operation of the speech synthesizer using the waveform editing method will be described with reference to FIG. In this case, the speech segment database (105) holds speech waveform data instead of the linear prediction coefficient vector. As a method of retaining a waveform, a method of performing zero-phase or minimum-phase processing on a waveform is widely used so that change control of a pitch frequency and a time length can be easily performed on voiced section data.
[0015]
Based on the speech unit information determined from the reading information, speech waveform data necessary for synthesis is extracted from the speech unit database (105). The amplifier ( 204 ) appropriately amplifies the amplitude of the speech waveform based on the voice volume (amplitude information) of the prosody information. When the speech unit is voiced, the waveform superposition unit ( 205 ) arranges the amplified speech waveform at appropriate time intervals based on the information of the pitch (pitch frequency) of the prosody information and performs windowed addition. Output. If the speech unit is unvoiced, the amplified speech waveform is output from the output terminal ( 206 ) as it is.
[0016]
[Problems to be solved by the invention]
In the system as exemplified in Japanese Patent Application Laid-Open No. 60-208799, only one set of linear prediction coefficients is held in accordance with a phoneme, and a temporal change of a spectrum included in a natural speech signal cannot be expressed. There was a problem that was insufficient. In order to solve this problem, a method of holding spectrum information for each analysis frame is generally used.
[0017]
Therefore, a conventional speech synthesizer holds a large amount of information in a speech unit database for the purpose of obtaining high-quality sound synthesis. For example, if a VCV segment is used as a synthesis unit, about 1000 types of segments are required, and the time length of one segment holds, for example, an information amount of about 200 ms on average. If one frame is 5 ms, it is 40 frames.
[0018]
Here, in the case of the synthesis method of the waveform editing method, since the waveform data is stored in the speech unit database, there is a problem that the data capacity is several MByte. When the parameter synthesis method is used, the time series of the linear prediction coefficient vector is stored in the speech unit database. For example, when the data of the linear prediction coefficient vector per frame is quantized and expressed by 40 bits, a data capacity of about 200 KByte is required only for the information amount of the linear prediction coefficient.
[0019]
In the field of audio compression, there is a method of reducing the amount of information on a linear prediction coefficient vector by using a moving average prediction and a vector quantization method (Proceedings of the Acoustical Society of Japan, October 5-19, 2-5-12). ). However, even in this case, there is a problem that the information amount of the linear prediction coefficient vector per frame is about 20 bits, and the information amount is still large.
[0020]
[Means for Solving the Problems]
This onset Ming, a text analyzer that converts the input string to read information only, and prosody generator for generating the prosodic information of the reading information is input voice of the height and the size and the utterance speed, to read information It is input comprises a voice segment selector for selectively outputting the speech unit information from the speech unit database, the speech synthesizer to create a synthesized speech based on the prosodic information and the speech segment information input , a speech synthesizer, the speech synthesizer, the reference by storing a plurality of reference codebook obtained by separately vector quantizing the speech feature vector of said reference when classified by reference in a plurality of phonemes A vector codebook and a speech feature vector selector for selecting a speech feature vector corresponding to the speech unit information from the reference-specific vector codebook based on the speech unit information selected by the speech unit selector, Equipment In which, in the speech unit database, and selection criteria information for selecting the reference codebook with is determined from the read information, and the index of the reference by the vector codebook, stored for each speech unit The speech unit information output from the speech unit selector includes selection reference information corresponding to the selected speech unit information and an index of a criterion vector codebook. The speech feature vector selector selects, from the reference codebook selected from the reference-specific vector codebook based on the selection reference information included in the input speech unit information, an index included in the speech unit information. It is characterized in that a voice feature vector is selected based on the selected voice feature vector .
[0021]
Further , at least one speech feature vector is stored in the reference codebook .
[0022]
Moreover, the selection criterion information is phoneme information, wherein the base Junfu Gocho is sound Motofu No. book.
[0023]
Alternatively, the selection criterion information is phoneme information, and the reference codebook is a plurality of category codebooks obtained by vector-quantizing a speech feature vector separately for each category as the reference, and the speech synthesizer includes: Based on the phoneme category table in which the category and the phonemes belonging to this category are stored in association with each other, and the selection criterion information included in the input speech unit information and the phoneme category table, the phoneme information Wherein the speech synthesizer selects a category codebook from the reference-specific vector codebook based on the selected category. And
[0024]
The front as Kion voice feature vectors, characterized by using the linear prediction coefficients.
[0025]
Alternatively, before the Kion voice feature vectors, characterized by using a voice waveform.
[0026]
In addition, the present invention is to convert the read only information the input string by text analysis means, before Symbol text analysis means is input to read information is converted by generating a prosodic information by the prosody generation means, before The read-out information is input, and speech unit information is selectively output from a speech unit database by a speech unit selection unit, and synthesized by a speech synthesis unit based on the input prosody information and the speech unit information. to create a sound, a speech synthesis method, by criterion storing a plurality of reference codebook obtained by separately vector quantizing the speech feature vector of said reference when classified by reference in a plurality of phonemes with use of the vector codebook, wherein the speech unit database, and selection criteria information for selecting the reference codebook with is determined from the read information, the reference by vector And the index of the codebook may be stored for each speech segment, by said speech synthesis means, criteria selected from the reference by vector codebook based on the selection criterion information included in the input speech segment information A speech feature vector is selected from the codebook based on an index included in the speech unit information .
[0027]
In addition, the present invention provides a computer including a text analyzer, a prosody generator, a speech unit selector, a speech synthesizer, and a speech feature vector selector in the speech synthesizer of the present invention .
A speech synthesis program that functions as a computer is recorded .
[0028]
BEST MODE FOR CARRYING OUT THE INVENTION
FIG. 1 shows an overall schematic configuration of a speech synthesizer of the present invention. The schematic configuration is the same as the conventional example. The present invention differs from the prior art in that the internal configuration of the speech synthesizer (106) and that the speech feature vector is compressed and stored in the speech unit database (105).
[0029]
FIG. 2 shows a portion of the speech synthesizer 106 extracted. FIG. 9 shows an example of the configuration of the entire text-to-speech synthesizing apparatus according to the present invention, combining FIG. 1 and FIG. Hereinafter, an embodiment of the speech synthesizer of the present invention will be described with reference to FIGS. The same components as those in FIG. 5 described in the description of the related art are given the same numbers. A component different from the prior art is a reference-based vector codebook (208) that inputs selection reference information determined from output information of a speech unit selector (104) and selects a vector codebook to be used from among a plurality. And a speech feature vector selector (207) for inputting an index determined from output information of the speech unit selector (104) and a reference-specific vector codebook (208) and selecting a speech feature vector. .
[0030]
As the selection criterion, for example, phoneme information, phoneme category information, or a criterion in which phonemes are classified by their acoustic characteristics (voiced, unvoiced, frictional, etc.) is used. As the speech feature vector, for example, data expressing speech features such as a linear prediction coefficient, a speech waveform, or a formant frequency and a bandwidth are used.
[0031]
For convenience of explanation, an example in which phoneme information is used as a selection criterion and a linear prediction coefficient is used as a speech feature vector will be described first. FIG. 3 shows an embodiment of a speech synthesizer using phoneme information as a selection criterion. The difference between the conventional example and the present application lies in the structure of the vector codebook (301). The phoneme-specific vector codebook (301) holds a codebook of speech feature vectors for each phoneme using phoneme information as a selection criterion. For example, a vector codebook is provided for each phoneme such as vowels such as a, i, u and consonants such as p, t, s. Here, the speech feature vector uses a linear prediction coefficient, and more specifically, a line spectrum pair or a cepstrum coefficient calculated from the linear prediction coefficient.
[0032]
Here, if a vector codebook of linear prediction coefficients is provided for each phoneme, those having similar spectral shapes are collected among the same phonemes. FIG. 7 shows an example in which 64 types of codebooks are created for each phoneme. For the phonemes a, i, u from the upper left side, and for the phonemes s, m, g from the lower left side, the first five speech feature vectors of the 64 types of codebooks are represented by a power spectrum. . It can be seen that the power spectrum of the consonant s has a very large variation, but the power spectrum of other phonemes has a similar shape. It can be understood that the feature vector can be efficiently expressed by vector-quantizing the feature vector for each phoneme.
[0033]
Although the overall configuration of the speech synthesizer is configured as shown in FIG. 1, the contents of the speech unit database (105) are different from the conventional one. As described in the related art, in the related art using the parameter synthesizing method, the time series information of the linear prediction coefficient vector was held, whereas in the technique of the present invention, the phoneme-specific vector code corresponding to the speech unit was used. Keeps a time series of book indexes. FIG. 8 schematically shows this state. This example shows a situation when "ito" is selected as an example of a VCV unit from the speech unit database (105). In the present embodiment, idx indicates an index indicating a codeword in a codebook obtained by vector-quantizing spectrum information. Between frame numbers 1 to seg1-1, the codebook of phoneme i in the phoneme-specific vector codebook (301) is selected, and the data of idx [1] to idx [seg1-1] is used to select the codebook of the vector codebook of i. The audio feature vector of the corresponding index is output from the audio feature vector selector (207). Similarly, between the frame numbers seg1 to seg2-1, the codebook of the phoneme t of the phoneme-based vector codebook (301) is stored, and between the frame numbers seg2 to seg3, the phoneme o of the phoneme-based vector codebook (301) is stored. A codebook is selected, and a similar process is performed.
[0034]
Unlike general speech compression, text speech synthesis has uttered character string information, so when compressing and recording speech feature vector information, phoneme information is obtained and vector quantization is performed for each phoneme. This makes it possible to utilize the fact that the spectrum information has a bias, and to efficiently compress the information.
[0035]
Here, the data amount is compared with the conventional method. In order to obtain a high-quality synthesized sound, for example, in order to quantize and express a 10th-order linear prediction coefficient, scalar quantization requires about 40 bits and vector quantization requires 20 bits or more. If the linear prediction coefficient is vector-quantized for each phoneme as shown above, the data has a very similar shape as the spectrum shape as exemplified above, so that it can be expressed with relatively little information. According to a simple subjective evaluation test of sound quality, good sound quality can be obtained even with an information amount of about 6 bits (64 types of spectrum patterns) for each phoneme, and significant information compression can be realized. Here, the codebook size of the vector quantization for each phoneme may be different for each phoneme. The operation after the speech feature vector is extracted and set in the synthesis filter (205) is the same as the conventional example.
[0036]
Next, an embodiment in which a vector codebook for each phoneme category is used will be described. In the above-described speech synthesizer based on phoneme information, it is necessary to hold a vector codebook for each phoneme. As exemplified above, for example, when a speech feature vector of 64 patterns is held for each phoneme, the number of phonemes is 32, and the number of phonemes is represented by 40 bits, the data amount of the phoneme-specific vector codebook is about 10 KByte. there were.
[0037]
Acoustically similar phonemes also have similar feature vectors, and having a codebook for each acoustically similar phoneme category rather than for each phoneme can reduce the capacity of the vector codebook.
[0038]
FIG. 4 shows a configuration for realizing this. In the configuration of FIG. 4, a phoneme category table (402) indicating which category each phoneme belongs to is provided, and the input phonemes are selected by the category selector (403) with reference to this category table. . The vector codebook of the speech feature vector is stored in the vector codebook for each phoneme category (401) for each phoneme category, and a codebook to be used is selected based on the selected category information. For example, if the input phoneme is d, since the phoneme d belongs to category 2, the phoneme category-specific vector codebook uses the second type codebook.
[0039]
By appropriately configuring phoneme categories having similar speech feature amounts, speech feature vectors can be held efficiently, and sound quality degradation can be kept small, and codebooks can be held more than phoneme-specific vector codebooks. The amount of memory required for this can be reduced. Subsequent operations are the same as in the embodiment of FIG.
[0040]
In the above description, the speech synthesizing apparatus of the parameter synthesizing method using the linear prediction coefficient as the audio feature vector has been described. However, a speech synthesizer based on the waveform superposition method may be used. In this case, a speech waveform is used as the speech feature vector. In particular, by performing zero-phase processing or minimum-phase processing and normalizing the amplitude, a vector codebook of a waveform can be configured based on a waveform distortion measure. Note that all of the above devices can be realized using a processing device such as a computer.
[0041]
【The invention's effect】
By using the vector codebook for each selection criterion (for example, phoneme) shown in the present invention, the speech feature vector can be efficiently compressed and expressed, and the capacity of the speech unit database can be greatly reduced.
[Brief description of the drawings]
FIG. 1 is a diagram illustrating a configuration example of an entire text-to-speech synthesis apparatus.
FIG. 2 is a diagram illustrating a configuration example of a speech synthesizer of the present invention.
FIG. 3 is a diagram illustrating a configuration example of a speech synthesizer of the present invention.
FIG. 4 is a diagram illustrating a configuration example of a speech synthesizer of the present invention.
FIG. 5 is a diagram illustrating a configuration example of a conventional speech synthesizer.
FIG. 6 is a diagram for explaining the operation of a conventional waveform superposition type synthesizer.
FIG. 7 is a diagram illustrating an example of data in a speech unit database.
FIG. 8 is a diagram illustrating an example of a speech feature vector codebook for each phoneme.
FIG. 9 is a diagram illustrating a configuration example of the entire text-to-speech synthesis apparatus of the present invention.
[Explanation of symbols]
101 input terminal 102 text analyzer 103 prosodic generator 104 speech unit selector 105 speech unit database 106 speech synthesizer 107 output terminal 201 pulse generator 202 noise generator 203 sound source switcher 204 amplifier 205 synthesis filter 206 output terminal 207 Speech feature vector selector 208 Vector codebook 301 by reference Vector codebook 401 by phoneme Vector codebook 402 by phoneme category Phoneme category table 403 Category selector 501 Inverse quantizer 502 Parameter quantization table 506 Speech synthesizer 601 Amplifier 602 Waveform Superimposer 603 output terminal

Claims (8)

And text analysis to convert the input string to read information only,
A prosody generator for generating prosody information is inputted to read information converted by the text analyzer,
Together with the read information is input, a speech unit selection unit for selectively outputting the speech unit information from the speech unit database,
A speech synthesizer comprising a speech synthesizer that creates a synthesized speech based on the input prosody information and the speech unit information,
The speech synthesizer includes: a reference-specific vector codebook storing a plurality of reference codebooks obtained by vector-quantizing a speech feature vector separately from the reference when the plurality of phonemes are classified based on a certain reference; A speech feature vector selector for selecting a speech feature vector corresponding to the speech unit information from the reference-specific vector codebook based on the speech unit information selected by the segment selector,
In the speech unit database , selection reference information determined from the reading information and for selecting the reference codebook and an index of the reference-specific vector codebook are stored for each speech unit. ,
The speech unit information output from the speech unit selector includes selection reference information corresponding to the selected speech unit information and an index of a reference-specific vector codebook,
The speech feature vector selector includes an index included in the speech unit information from a reference codebook selected from the reference-specific vector codebooks based on the selection reference information included in the input speech unit information. A speech synthesizing apparatus characterized in that a speech feature vector is selected on the basis of the speech feature vector . The speech synthesizer according to claim 1, wherein at least one speech feature vector is stored in the reference codebook . The selection criterion information is phoneme information,
The group Junfu Gocho speech synthesis apparatus according to claim 1 or 2, characterized in that a sound Motofu No. book. The selection criterion information is phoneme information,
The reference codebook is a plurality of category codebooks obtained by vector quantization of speech feature vectors separately for each category as the reference,
The speech synthesizer,
A phoneme category table in which the category and phonemes belonging to this category are stored in association with each other;
A category selector for selecting a category corresponding to the phoneme information based on the selection criterion information included in the input speech unit information and the phoneme category table
Having
The speech synthesizer selects a category codebook from the reference-specific vector codebook based on the selected category.
Speech synthesis apparatus according to claim 1 or 2, characterized in that it is so. Before the Kion voice feature vectors, the speech synthesis apparatus according to any one of claims 1 to 4, characterized by using the linear prediction coefficients. Before the Kion voice feature vectors, the speech synthesis apparatus according to any one of claims 1 to 4, characterized in that a speech waveform. Converts to read only information the input string by text analysis means, before Symbol text analysis means is input to read information is converted by generating a prosodic information by the prosody generation means, before Symbol reading information is input select output speech unit information from the speech unit database by speech unit selection means Te, to create a synthesized voice by the voice synthesizing means based on the prosody information input and said speech unit information, a speech synthesis method,
Using a reference-specific vector codebook that stores a plurality of reference codebooks obtained by vector-quantizing a speech feature vector separately from the reference when a plurality of phonemes are classified based on a certain reference,
In the speech unit database , selection reference information determined from the reading information and for selecting the reference codebook and an index of the reference-specific vector codebook are stored for each speech unit. ,
From the reference codebook selected from the reference-specific vector codebooks based on the selection reference information included in the input speech unit information, by the speech synthesis unit, based on the index included in the speech unit information. A speech synthesis method characterized by selecting a speech feature vector . Computer
A text analyzer, a prosody generator, a speech unit selector, a speech synthesizer, and a speech feature vector selector according to claim 1 .
A computer-readable recording medium having recorded thereon a speech synthesis program functioning as a computer .

Images