JPH03189697A - Regular voice synthesizing device - Google Patents

Abstract

PURPOSE:To improve the working efficiency of an input and rewriting of a document, etc., and also, to obtain a regular voice synthesizing device having a high tone quality obtaining a more complete composite voice by synthesizing a prescribed parameter, a pitch pattern and amplitude and obtaining the composite voice. CONSTITUTION:A document into which phoneme information obtained by a syntex analysis and a symantic analysis of an arbitrary document is inserted as a phoneme symbol S is inputted from a keyboard, etc., and its data is trans mitted to a phoneme parameter converting circuit 4. To the circuit 4, an output from a phoneme parameter file 3 in which a list of parameters is stored is also supplied, therefore, the data of the symbol S is converted to a phoneme parameter, based on this output. Also, a signal from a microphone 2 passes through an A/D converter 9 and an output signal is transmitted to a pitch pattern determining circuit 5 and an amplitude determining circuit 6, and by the circuit 5, 6, a pitch pattern and amplitude of a voice based on the output signal are determined. In a synthesizing 7, the phoneme parameter, the pitch pattern and the amplitude synthesize a voice, and this composite voice output is outputted from an output terminal 8.

Description

[Detailed description of the invention] [Industrial application field] The present invention generates speech using a so-called rule synthesis method.

[Summary of the invention]

The present invention improves work efficiency by converting phonological information into predetermined parameters, converting speech into electrical signals, determining the pitch pattern and amplitude of the signals, and synthesizing them to obtain synthetic speech. The present invention provides a regular speech synthesizer that is capable of obtaining high-quality synthesized speech and more complete synthesized speech.

[Conventional technology]

Artificially creating sounds is called speech synthesis, and this method of speech synthesis includes recording editing method, parameter editing method,
It can be classified into rule synthesis methods, etc.

The recording/editing method described above stores (records) voices uttered by a person in units of words, phrases, etc., and reads them out and aggregates them as needed to synthesize voices. It is.

The above parameter editing method uses a single word, phrase, etc. as a unit, as in the case of the above recording editing method, but it analyzes the voice uttered by a person in advance based on a speech generation model and stores it in the form of a parameter time series. This method uses connected parameter time series as necessary to drive a speech synthesizer to synthesize speech. Specific speech synthesis devices include a so-called channel vocoder, LSP (line spectrum pair), and PA.
A synthesis device based on a linear predictive analysis method such as the RCOR (partial autocorrelation) method is used.

The rule synthesis method described above stores relatively short audio parameters as units of synthesis, and then manually adds settings such as connection methods, length adjustments, accents, and intonation, and controls the synthesizer using rules. Generate parameters. Then, by giving the generated voice parameters to a synthesizer, a synthesized sound is artificially created. Using this method, any text can be synthesized based on given character information. The input phonetic symbols such as phonetic symbols and prosodic symbols such as accents are converted from the input symbol series into control parameters using a list of parameters, a list of accent patterns, etc. corresponding to each phoneme. Further, parameters are modified in consideration of tonal combination, etc., according to rules.

Here, the principles of generating speech using the above-mentioned rule synthesis method include the so-called vocal tract analog, which electrically simulates the propagation of sound waves in the vocal tract, and the frequency spectrum structure of speech as a result of articulation, that is, resonance and anti-resonance. Terminal analog, which electrically simulates only the phenomenon, or speech synthesis based on linear predictive analysis is used. In some cases, human voice is used as the material to extract the characteristic parameters of the basic unit, and in some cases it is completely artificial. Phonemes are the basic unit of 0 synthesis, and only 30 to 50 types are needed, so the storage capacity is small, but the combination rules are quite complex, and it is difficult to obtain good quality. For,
A slightly larger unit is often used. In the case of Japanese, there are 100 syllables (CV unit; C
is a consonant and ■ is a vowel) is often used. In order to obtain better synthesized speech, a method using CVC as a unit is also being considered. However, if all possible CVCs in Japanese are prepared, the number will be enormous, 5,000 to 6,000, so approximately 1,000 types of CVCs, CVs, and VC units, which are frequently used, are used. A method using ■C■CV is also being considered, and in this case the number of units would be 700 to 800.

FIG. 4 shows a schematic configuration of a conventional regular speech synthesis device.
In FIG. 4, for example, a document in which phonological information and prosodic information are obtained by syntactic analysis and semantic analysis of an arbitrary document, and this phonological information and prosodic information are inserted as phonological symbol S and prosodic symbol R is a keyboard. It is input from etc. Above phonetic symbol S
The data is transmitted to the phoneme parameter conversion circuit 104. This phoneme parameter conversion circuit 104 is supplied with an output from a phoneme parameter file 103 which is made up of, for example, a memory and stores a list of parameters corresponding to phonemes. Therefore, the data of the phoneme symbol S is converted into phoneme parameters based on the output of the phoneme parameter file 103. This phonological parameter is sent to the synthesizer 107.

Further, the data of the prosodic symbol R is transmitted to the pitch pattern determining circuit 105 and the amplitude determining circuit 106. The pitch pattern determining circuit 105 and the amplitude determining circuit 106 determine the pitch pattern and amplitude of the speech based on the prosodic symbol R.

This pitch pattern and amplitude are transmitted to the synthesizer 107. The synthesizer 107 synthesizes speech from the above phoneme parameters, pitch pattern, and amplitude, and this synthesized speech output is output from the output terminal 108.

[Problem to be solved by the invention]

However, in the above-mentioned regular speech synthesis device, all inputs such as text, phonemes, prosodic symbols, etc. are inputted using a keyboard or the like. Therefore, these keyboard inputs take time,
Furthermore, since the metrical symbols indicate the position of the accent voice, etc., it is difficult to input them using a keyboard. Furthermore, when listening to the actually synthesized and output speech, the prosody may be incorrect. In such cases, it is necessary to rewrite the input prosodic symbols, so conventionally,
The above keyboard was used to directly delete and change prosodic symbols. As described above, in regular speech synthesis, inputting prosody and phonetic symbols, rewriting prosodic symbols, etc. takes a very long time, resulting in low work efficiency and often resulting in incomplete synthesized speech.

Therefore, the present invention has been proposed in view of the above-mentioned actual situation, and is a method of regular speech that can improve the work efficiency of inputting documents and rewriting prosodic information, and can obtain more complete synthesized speech. The purpose is to provide a synthesis device.

[i! ! Means for Solving the Problem] The regular speech synthesis device of the present invention has been proposed to achieve the above-mentioned object, and includes a parameter conversion means for converting phonological information into predetermined parameters, and a method for converting speech into electrical signals. comprising an acousto-electric conversion means for converting into a signal, a pitch pattern determination means for determining the pitch pattern of the output of the acousto-electric conversion means, and an amplitude determination means for determining the amplitude of the output of the acousto-electric conversion means, The predetermined parameter, the pitch pattern, and the amplitude are synthesized to obtain a synthesized speech.

[Effect]

According to the present invention, since the pitch pattern and amplitude as prosodic information are directly obtained from the output of the electro-acoustic conversion means that converts speech into an electrical signal, it is not necessary to input the prosodic information as prosodic symbols on the keyboard. do not have.

〔Example〕

Embodiments in which the present invention is applied will be described below with reference to the drawings.

FIG. 1 shows a schematic configuration of a regular speech synthesizer according to an embodiment of the present invention.

This regular speech synthesis device includes a phonological parameter conversion circuit 4 that converts phonological information into phonological parameters as predetermined parameters, a microphone 2 that is an acoustic-to-electrical conversion means that converts speech into an electrical signal, and Pitch pattern determination circuit 5 for determining the output pitch pattern
and an amplitude determining circuit 6 for determining the amplitude of the output of the microphone 2, and a synthesizer 7 synthesizes the phonological parameters, the pinch pattern, and the amplitude to obtain synthesized speech. This is what I did.

That is, in FIG. 1, for example, phonological information can be obtained by syntactically and semantically analyzing an arbitrary document, and a document with this phonological information inserted as a phonological symbol S can be input from a keyboard or the like. ing. The data of the phoneme symbol S is transmitted to the phoneme parameter conversion circuit 4. The phoneme parameter conversion circuit 4 is also supplied with an output from a phoneme parameter file 3 which is made up of, for example, a memory and stores a list of parameters corresponding to phonemes. Therefore, the data of the phoneme symbol S is converted into phoneme parameters based on the output of this phoneme parameter file 3. This phoneme parameter is sent to the synthesizer 7.

Further, in this embodiment, the pitch pattern and amplitude as prosodic information are directly determined from the signal obtained from the microphone 2 via the analog/digital (A/D) converter 9. That is, the A/D converter 9
The output signal is transmitted to the pitch pattern determining circuit 5 and the amplitude determining circuit 6. The pitch pattern determining circuit 5
．． The amplitude determining circuit 6 determines the pitch pattern and amplitude of the voice based on the output signal, and the pitch pattern and amplitude are transmitted to the synthesizing device 7. The synthesizer 7 synthesizes speech from the above-mentioned phoneme parameters, pitch pattern, and amplitude, and this synthesized speech output is outputted from the output terminal 8.

Here, the pitch pattern determining circuit 5 is configured to, for example,
A configuration as shown in the figure can be used.

The configuration shown in FIG. 2 determines the pitch by using a so-called autocorrelation method. In FIG. 2, the output signal from the A/D converter 9, that is, the audio signal, is
The signal is supplied via an input terminal 11 to a prediction residual filter 12 composed of, for example, a predictor and an adder (or subtracter). The audio signal is also supplied to a linear predictive analysis (LPG analysis) circuit 13. The linear predictive analysis circuit 13 performs linear predictive analysis on the audio signal every 20 m5, for example. The LPG coefficients from the linear prediction analysis circuit 13 are transmitted to the prediction residual filter 12. Therefore, in the prediction residual filter 12, for example, a prediction signal from a predictor based on the LPG coefficients is sent to an adder as a subtraction signal, and the prediction signal is subtracted from the audio signal to generate a prediction error signal. is obtained. This prediction error signal is transmitted to an analysis window multiplication circuit 14 that multiplies it by an analysis window (for example, a Hamming window) having a length of, for example, about 30 m5.

The output of this analysis window multiplication circuit 14 is transmitted to an autocorrelation coefficient calculation circuit 15, and the autocorrelation coefficient calculation circuit 15 calculates
Autocorrelation function r(n)+n・0,1. ...is required. Thereafter, this autocorrelation function r(n·) is sent to the maximum value search circuit 16, where the autocorrelation function r(n.)
The maximum value of (n) is determined. This autocorrelation function r
The pitch is the maximum value (n) and is output from the output terminal 17. In other words, the above process is carried out for example on 20m5.
By repeating each time, the pitch pattern can be found. In addition to the autocorrelation method described above, there are
For example, it is also possible to use the cepstral method.

Here, the cepstrum is defined as the inverse Fourier transform of the logarithm of the short-time amplitude spectrum of a waveform, and has the characteristic that the spectral envelope and fine structure can be approximately separated and extracted.

Further, in the amplitude determination circuit 6 of FIG. 1, the amplitude is determined from the audio signal by the following processing.

That is, the A/D transmitted to the amplitude determining circuit 6
The audio signal from the converter 9 is subjected to linear prediction analysis in the amplitude determination circuit 6, for example, every 20 m5, and a prediction residual signal is formed. Thereafter, by calculating the square root of the average power of the prediction residual signal for each section of 20 m5, this can be obtained as the amplitude.

Furthermore, in the apparatus shown in FIG. 1, the temporal correspondence between phoneme parameters, pinch patterns, and amplitudes can be performed as follows. Here, the third a indicates the phonological parameter sequence, b in Fig. 3 indicates the pinch pattern of the input speech signal, and C in Fig. 3 indicates the amplitude pattern, that is, as shown in a in Fig. 3. The time length T of the phoneme parameter string is fixed in advance for each phoneme parameter string. In a of FIG. 3, tl is the time indicating the starting point of the target phoneme parameter string, and t! is the time indicating the end point of the target phonetic parameter sequence. Also, b and C in Figure 3
As shown in the figure, the starting point time [, ~ ending point time t4 of the input audio signal is determined by the fact that the short-time power of the input audio signal is greater than or equal to a certain value. Therefore, in order to temporally associate phonetic parameters, pitch patterns, and amplitudes, it is necessary to create a pitch pattern so that the time length T2 from time t to time t4 is equal to the time length T. and expand/contract the amplitude pattern linearly on the time axis. By doing as described above, the starting point time 1.
It is possible to make the end point time t2 and the start point time t to the end point time t4 of the pitch pattern and the amplitude pattern coincide with each other.

Here, as an example of regular speech synthesis, if we take the document "This is Japan," as an example, the phonetic symbol S input from the keyboard is "Kokohanihondesu."
becomes. Conventionally, the prosodic symbol R representing prosodic information in the phonetic symbol S of "Kokohanihondesu" includes, for example, a symbol indicating an accent (for example, A) and a symbol indicating a pause (
For example, P) is inserted. Therefore, by adding the above-mentioned phonetic symbol S and metrical symbol R to the document "This is Japan," it becomes "Kokoha P Niho Ande As P."
In contrast, in this embodiment, first, "Kokohanihondesu" is input as the phonetic symbol S, and the voice "This is Japan" is input directly into the microphone 2.
There is no need to use the prosodic symbol R, that is, there is no need to input the prosodic symbol R using a keyboard, etc., and conventionally, if the prosody is incorrect, the prosody symbol R can be rewritten using the keyboard. On the other hand, in this embodiment, since the voice is input directly from the microphone, there is no need for rewriting using the keyboard. Therefore, with the regular speech synthesis device of this embodiment, it is possible to input and correct documents in a very short time, improving work efficiency and making it possible to obtain more complete synthesized speech.

By the way, in regular speech synthesis, in order to improve the quality of synthesized speech, there is a method in which long syllables (including multiple musical pieces) are used as synthesis units. The regular speech synthesis device in this case is equipped with the above-mentioned syllable file, and this syllable file has synthesis units as a syllable parameter string analyzed from continuous speech. As the above synthetic unit, the above-mentioned C■
(consonant vowel), VCV (vowel-consonant-vowel), CVC (
consonant-vowel-consonant), etc. In this case, the regular speech synthesizer sequentially calls out the necessary synthesis units based on the phonetic symbol string, and connects each synthesis unit using signal processing so that the voice becomes close to the real voice. In general, the synthesized sound quality tends to be better as the synthesis unit is longer. However, the longer the synthesis unit is, the larger the capacity required for the syllable file becomes, which becomes a major bottleneck in actually creating a synthesis device.

For this reason, it is conceivable to create a syllable file by compressing and encoding the syllable parameters. Therefore, when synthesizing speech, compression codes for necessary syllable parameters are extracted from the syllable file, decoded, and then used.

Here, examples of the above-mentioned compression encoding include so-called vector quantization.

Further, the phonological parameters specifically represent the spectral envelope of speech, and for example, the so-called PAR
parameters by COR analysis.

Examples include linear prediction filter coefficients (LPG coefficients), line spectrum pair (LSP) coefficients, cepstral coefficients, and mel cepstral coefficients.

In other words, by compressing and encoding syllable parameters to create a syllable file as described above, the storage capacity required for the syllable file is reduced, making it possible to realize a high-quality speech synthesizer that uses long syllables as synthesis units. become.

Further, the present invention can be applied to a so-called text synthesis method in addition to the above-described rule synthesis. Here, the above-mentioned text synthesis method is a speech synthesis method that generates a sequence of control parameters according to rules from text written in normal calligraphy. In this text synthesis method, text is first broken down into words and smaller morphemes, and after performing language-level processing such as syntactic analysis and semantic analysis, it is converted into phonological symbols using a dictionary, accents are added, etc. Performs audio level processing. In the case of Japanese, since the input sequence includes kanji, a kanji phonetic dictionary is required. For phonograms such as English, a phonological conversion dictionary is required for every word, but if you consider all the changes in verb tense, person, and adjective, the dictionary becomes enormous. Therefore, words are broken down into morphemes, which are smaller units such as roots, prefixes, and suffixes, and these are converted into phoneme sequences using rules. In this way, text synthesis is a method of automatically synthesizing speech by a machine, similar to how a human reads aloud text written in ordinary letters. It can be said that this is the final form of speech synthesis that goes all the way to the top. Parameter editing methods and rule synthesis methods are part of text synthesis methods, and the techniques used here are also used in text synthesis.

〔Effect of the invention〕

The regular speech synthesis device for speech data of the present invention converts phonological information into predetermined parameters, converts speech into electrical signals, determines the pitch pattern and amplitude of the signals, and synthesizes them to produce synthesized speech. By making it possible to obtain
The work efficiency of inputting and rewriting documents, etc. is improved, and
It becomes possible to obtain more complete synthesized speech.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing a schematic configuration of an apparatus according to an embodiment of the present invention, FIG. 2 is a block diagram showing a configuration for determining pitch patterns, and FIG. 3 shows temporal correspondence between phonetic parameters, pitch patterns, and amplitudes. FIG. 4 is a block diagram showing a schematic configuration of a conventional device.・・・・・・・・・・・・Microphone・・・・・・
・・・・・・Phonological parameter file・・・・・・・・・
...Phonological parameter conversion circuit...
・Pitch pattern determination circuit・・・・・・・・・Amplitude determination circuit・・・・・・・・・Synthesizing device

Claims (1)

[Scope of Claims] Parameter converting means for converting phonetic information into predetermined parameters; acoustic-to-electrical converting means for converting speech into electrical signals; and a pitch pattern for determining the pitch pattern of the output of the acoustic-to-electrical converting means. and an amplitude determining means for determining the amplitude of the output of the acousto-electric conversion means, and is characterized in that the predetermined parameter, the pitch pattern, and the amplitude are synthesized to obtain a synthesized speech. A rule-based speech synthesizer.