JPH01197799A - Articulation and sound source parameter producing method for sound synthesizer - Google Patents

Abstract

PURPOSE:To obtain articulation and sound source parameters which are close to actual sounds by finding the sound source parameters from the response data of the input side of a digital filter having an optionally set articulating parameter when actual sound data are given from the output side of the filter. CONSTITUTION:When a sound source (syllable or phoneme) is commanded, a control section 16 designates the sound source to a sound reading-out section 12 and, at the same time, sets sectional-area factor data corresponding to the sound source in a digital filter calculating section 13 and registers the data in a parameter registering section 15. Upon receiving the designation of the sound source, the section 12 reads out the sampling data of the sound source and successively gives the data to the output side of the calculating section 13. The section 13 performs calculation in accordance with a calculation formula decided from an equivalent circuit and the set sectional-area factor and outputs sampling data having the waveform of the sound source against the waveform of the inputted sounds. A parameter calculating section 14 registers a sound source parameter in the parameter registering section 15 against the data.

Description

DETAILED DESCRIPTION OF THE INVENTION A. Field of Industrial Application The present invention relates to a speech synthesis device using a rule synthesis method, and particularly to a method for generating articulatory parameters and sound source parameters.

B1 Summary of the Invention The present invention provides a method for generating articulatory parameters and sound source parameters associated with syllables or phonemes, when real speech data is given from the output side of a digital filter having arbitrarily set articulatory parameters. By determining the sound source parameters from the side response data and using them in combination with the articulatory parameters at this time as parameters for synthesis of the syllable or phoneme, it is possible to improve speech quality while simplifying the digital filter and sound source parameter generation. This is something that can be improved.

C0 A speech synthesis apparatus using a conventional technical rule synthesis method has a configuration shown in FIG. 3, for example. The text analysis unit 1 performs text analysis on a character string of a Japanese input text using a dictionary 1a and a text analysis device tb. In addition to a dictionary for converting the pronunciation of words, the dictionary 1a has a Japanese grammar dictionary for dividing words into clauses and phrases, and also has a dictionary of rules for word accents and basic intonation. Sentence analysis device! b
converts the input text into a phoneme symbol string of phonemes or syllables with reference to the dictionary 1a, and also generates prosodic information such as word accents and basic intonation.

The speech synthesis rule section 2 is composed of a file 2a and a parameter generation device 2b. The file 2a stores feature parameters of phoneme units, their connection rules, and prosodic information control rules. The parameter generation device 2b generates a control parameter sequence in which feature parameters for phoneme information are linked together with information such as its duration, and also generates a sound source pattern sequence in which pitch, energy, and intonation processing of the sound source is processed using prosody information.

The speech generation section 3 is composed of a sound source generation device 3a, a speech synthesis digital filter 3b, and a speech converter 3c. The sound source generation device 3a generates sound source signals such as pitch and energy according to the sound source pattern sequence. In the digital filter 3b, the parameters of the Percoll coefficient, transfer function, or formant frequency are changed according to the control parameter string, and a synthesized speech data string is obtained as a response output to the sound source signal with these parameters. The audio converter 3c is a filter 3b
The output of the converter is converted into an analog signal to obtain an audio waveform, and synthesized audio is output by an electro-sound conversion means such as a speaker.

In the speech synthesis device as described above, generation of a speech waveform is separated into sound source generation and articulation based on vocal tract shape, and the sound source and articulation are modeled in an electrical equivalent circuit. For example, a vocal cord vibration model using a two-mass model is required for a sound source, and a vocal tract cross-sectional area model and an articulation model that directly determines the structure of the articulatory organ and its motion characteristics are required for articulation. Speech characteristic parameters are determined by such modeling, but instead of directly handling the voice waveform, the voice characteristic parameters are determined by analyzing the voice spectrum. For example, the short-time spectrum of speech is decomposed into a spectral envelope that changes slowly with frequency and a spectral fine structure that changes finely periodically (voiced sounds) or aperiodically (unvoiced sounds). The spectral envelope corresponds to the overall resonance and anti-resonance characteristics of the vocal organs, and the spectral fine structure corresponds to the periodicity of the sound source. To extract the spectral envelope, a nonparametric analysis method using autocorrelation function analysis or cepstral analysis, or a parametric analysis method using ABS or linear predictive analysis (Percoll analysis, LSP analysis, etc.) is used.

Also, among the spectral fine structures, the fundamental frequency of the sound source (
Correlation processing and spectral processing methods are used to extract pitch).

In addition, to find the characteristic parameters of articulation, we consider the change in the cross-sectional area of the vocal tract as a cascade of many one-dimensional acoustic tubes, and express it as a linear combination of forward traveling waves and backward traveling waves in each acoustic tube. Modeling is performed using ratios and reflection coefficients, and modeling using resonance frequencies (formant frequencies) and amplitude spectrum characteristics. Furthermore, a model that directly represents the structure of the articulatory organ and its motion characteristics will be created.

D1 Problem to be Solved by the Invention Conventionally, various methods for determining voice characteristic parameters, as typified by linear predictive analysis, have attempted to efficiently and accurately determine the characteristics of voice waveforms and spectra using a small number of parameters. However, the synthesized sound that is actually output has less clarity than real speech.

It becomes significantly less natural. This is for example ABS
The method and linear predictive analysis method use feedback control to find parameters that minimize the error (prediction residual) between the value generated by the signal generation model and the observed value. This is because there are large differences between vocal fold vibration and vocal tract shape that change, and it is impossible to perform complete analysis and synthesis.

Furthermore, conventional methods require feedback control many times to obtain parameters, resulting in a problem of an extremely large amount of information processing. In this problem, the more detailed the modeling is, the more the amount of information processing increases, and the required storage capacity increases.

An object of the present invention is to provide an articulatory/sound source parameter generation method that can obtain articulatory parameters and sound source parameters that approximate actual speech while reducing the amount of information processing.

60 Means for Solving the Problems In order to achieve the above object, the present invention separates sound waveform generation into sound source generation and articulation based on vocal tract cross-sectional area, generates a sound source waveform based on sound source parameters, and generates voice based on articulatory parameters. In a speech synthesizer using a regular synthesis method, which applies a sound source waveform to a digital filter with a cross-sectional area of the vocal tract and obtains a speech signal as the output of the filter, the vocal tract of the digital filter is When the cross-sectional area coefficient is fixed and a real syllable or real phoneme waveform is given to the output side of the filter, a sound source parameter is determined from the signal waveform appearing on the input side of the filter, and the sound source parameter and the articulatory parameter are converted to the corresponding syllable. Or, it is characterized in that it is a parameter of a phoneme.

When adopting a vocal tract cross-sectional area function as an F1 effect digital filter, contrary to finding the response output to its input,
The manual waveform can be determined from the output waveform. This will be explained in detail below.

Articulation by the vocal tract can be viewed as a vibration phenomenon caused by traveling waves and reflected waves in a one-dimensional acoustic tube, which is a cascade of multiple acoustic tubes with different cross-sectional areas relative to the sound source waveform. It is appropriate to treat the acoustic tube as a distributed constant system due to the relationship between the length and thickness of the vocal tract and the sound wave length, and the propagation loss of the sound source can be ignored.

From the above, the vocal tract is represented by the equivalent model shown in Figure 2 (A) in which n one-dimensional acoustic tubes with cross-sectional areas A, ~An are connected in cascade, and each acoustic tube S, -Sn is represented by the LC component only. When considered as lossless distributed constant lines T, ~Tn, it is converted to the electric circuit shown in the same figure (a), and the output waveform when an impulsive voltage or current signal instead of a sound source is applied to this electric circuit is the audio waveform. It can be considered as a response output corresponding to . and,
To change the articulatory state of the vocal tract, the constants of each line T, ~T,, etc. are required.

By changing C completely, the surge impedance will be changed.

In the equivalent circuit of the same figure (a), when an impulsive current is applied in place of the sound source, each line T.

~Trl each has a current source whose internal resistance is the surge impedance Z1~Z, and is converted into an equivalent circuit with reflection and transmission at the connection point of each line T I""' T-.

This equivalent circuit is shown in the same figure. In the figure, E is the voltage source voltage corresponding to the sound source, and Z is the voltage source voltage corresponding to the sound source.

is its output impedance, ZL is the radiation impedance from the lips, i OA-i (1%-11A is the line T.

~Forward wave current at each connection point of Tn, foe~Ln-++a
is the backward wave current at each connection point, and Z I to Zn are the cross-sectional areas A, respectively.
From t~An, air density ρ, and sound speed C, ρC/Al~ρC/
Surge impedance that becomes An, IOA ~ 1111-
+1A and IIB-1nB are the current source currents & appearing at the connection point of the lines T and -Tn. 6~a (n-+lA and ale~ans are the shunt currents of the current source.

These currents have the relationship shown in FIG. 2 (1). In the same formula, S + a = A + / (A ++ A t), S
1A=At/(A++At)~S +n-+1B=A
(n-117(A+n-t++A11), S +1l
-119=A n/A (indicates n-11+A n).

Furthermore, the memory item indicates an item that stores the calculation result l steps before the current value. To calculate such a relational expression, give an initial value 0 to the memory item, and calculate the voltage E for each memory item sequentially up to the line T, up to T7 at the fundamental cycle, and calculate the final stage current 1ne at the fundamental cycle. , this current la
The articulatory speech output data □ can be obtained in g.

Here, the equivalent circuit in Fig. 2 (2) becomes a linear circuit, and the superposition principle is established, and when voltage E is set to zero and voltage E+ is applied in series to radiation impedance ZL, find the current at each part, That is, impedance E. The current can be found. Therefore, in the present invention, the vocal tract cross-sectional area coefficient of a digital filter is arbitrarily set in association with a syllable or a phoneme, and the signal waveform appearing on the input side when a real syllable or real phoneme is given to the output side of the digital filter. The sound source parameters of the syllable or phoneme are determined by analyzing this signal waveform, and the vocal tract cross-sectional area coefficient is also determined as an articulatory parameter.

The sound source parameters and articulation parameters obtained in this way can bring the output speech waveform closer to the actual speech while leaving the errors between the digital filter and the human vocal organs and the errors between the sound source waveform and vocal cord vibration unchanged. Further, in order to obtain the sound source parameters and articulatory parameters, unlike the conventional feedback method, it is only necessary to obtain a response to one actual voice, thereby reducing the amount of information processing required. Furthermore, the articulatory parameters make it unnecessary to model the vocal tract in detail and reduce the amount of information required as syllable or phoneme parameters, that is, it makes it possible to compress information and improve the calculation speed of digital filters or reduce the amount of calculations.

G. Embodiment FIG. 1 is a block diagram showing an embodiment of the present invention. In the voice registration section 11, an actual voice waveform in units of syllables or phonemes uttered by an announcer or the like is registered as sampling data. The voice succession section 12 reads sampling data of a designated syllable or phoneme from the voice registration section 11, and amplifies and outputs a voltage waveform proportional to this data. The digital filter calculation unit 13 receives the voltage waveform from the voice succession unit 12 as an input on the output side, sets the vocal tract cross-sectional area coefficient of the syllable or phoneme corresponding to the voltage waveform, and inputs the vocal tract cross-sectional area coefficient with this vocal tract cross-sectional area coefficient. Calculate the side response output (sound source data). The parameter calculation unit 14 analyzes the sound source data from the digital filter calculation unit 13 to obtain sound source parameters, and registers the sound source parameters in the parameter registration unit 15. In this registration, the vocal tract cross-sectional area coefficient set in the digital filter calculation section 13 is also registered. Control part 1
Reference numeral 6 controls each section according to commands from an input device 17 such as a keyboard, and displays the control status on a display device 18.

The articulation/sound source parameter generation process in the above configuration will be described in detail below.

When a sound source (syllable or phoneme) to be registered is instructed, the control unit 16 specifies the sound source to the audio reading unit 12, and sets and parameters the cross-sectional area coefficient data corresponding to the sound source in the digital filter calculation unit 13. Registration section 15
Register. The audio successive section 12 to which the sound source is specified reads the sampling data of the sound source from the audio registration section 11,
This data is sequentially applied to the output side of the digital filter calculation section 13. The digital filter calculation unit 13 performs calculations based on the calculation formula (the diagrammatic relational formula) determined from the equivalent circuit and the set cross-sectional area coefficient shown in FIG. 12, output impedance Z. Find the data string of voltage values as the response output. The data string obtained at this time becomes sampling data of the sound source waveform for the input voice waveform at the set cross-sectional area coefficient. The parameter calculating section 14 calculates sound source parameters from this data by analyzing the spectrum envelope, spectral fine structure, etc., and registers the sound source parameters in the parameter registration section 15 .

H. Effects of the Invention According to the articulatory/sound source parameter generation method of the present invention, the sound source waveform can be obtained for the sound source parameter of the syllable or phoneme to be determined by a single calculation by setting the vocal tract cross-sectional area function. This reduces the amount of information to be processed and the processing time compared to repetitive calculations using conventional feedback methods. Furthermore, although there is an error between the vocal tract cross-sectional area function and vocal tract equivalent circuit and the actual human vocal tract structure, this error is also included in the desired sound source waveform, and the sound source is used during actual speech synthesis. The error included in the waveform is canceled out by the error in the vocal tract cross-sectional area function of the digital filter, and only the error between the obtained sound source waveform and the sound source parameters is generated during speech synthesis. This has the effect of making synthesized speech closer to real speech, as well as simplifying the vocal tract cross-sectional area coefficient (for example, reducing the number of cascaded acoustic tubes) and simplifying the computation of cross-sectional area changes that vary complexly in each acoustic tube. The effect of simplifying processing and improving processing speed can be obtained.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing an embodiment of the present invention, Fig. 2 is an equivalent model of an acoustic tube and a calculation mode diagram for explaining the principle of the present invention, and Fig. 3 is a speech synthesizer using a regular synthesis method. FIG. 11... Voice registration section, 12... Voice reading section, 13.
...Digital filter calculation section, 14...Parameter calculation section,! 5... Parameter registration section, 16... Control section, 17... Input device, 18... Display device. 11+1-197799 (7)

Claims (1)

[Claims] (1) Separate sound waveform generation into sound source generation and articulation based on the vocal tract cross-sectional area, generate the sound source waveform using the sound source parameters, apply the sound source waveform to a digital filter for the vocal tract cross-sectional area based on the articulatory parameters, and then In a speech synthesizer using a regular synthesis method that obtains a speech signal as an output, the vocal tract cross-sectional area coefficient of the digital filter is fixed by an articulation parameter arbitrarily set in association with a syllable or phoneme, and a real syllable is output on the output side of the filter. Or, a speech synthesis device characterized in that a sound source parameter is determined from a signal waveform appearing on the input side of the filter when an actual phoneme waveform is given, and the sound source parameter and the articulation parameter are used as parameters of a corresponding syllable or phoneme. Articulatory/sound source parameter generation method.