JPS61259300A - Voice synthesization system - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] Like-Station! The present invention relates to a speech synthesis method for outputting synthesized speech from an arbitrary character arrangement, and more particularly to a spectral parameter generation method for obtaining high-quality synthesized speech.

Conventionally, in speech synthesis methods that can create any speech, phonemes,
The basic unit of synthesis is a syllable, VCV (vowel-consonant-vowel), etc., and the spectral parameters of each element and driving sound source signal are combined based on certain rules and fed to a speech synthesizer.
I'm trying to get a synthesized sound. Spectral parameters include LPG (Linear Predictive Coding) based on linear predictive analysis, PARCOR (Partial Autocorrelation), and LSP (Line Spectrum Pair), and the driving sound sources include impulse trains, white noise, Alternatively, the mainstream method is to use residual waveforms to improve synthesized sound quality.

Traditionally, the synthesis unit speech segment is a single syllable or word.

Spectral parameters have been created by cutting out the necessary parts of the human voice uttered as sentences and performing linear predictive analysis.Linear predictive analysis analyzes the sound source characteristics and sound path characteristics during the speech generation process.

Since all radiation characteristics are combined into one all-pole model, the parameters obtained by this include the sound source characteristics at the time of recording the elemental fragments, and are The drawback is that the sound quality deteriorates when the pieces are connected to obtain a synthesized sound.

The present invention was made in view of the above-mentioned circumstances.
In particular, this was done with the aim of obtaining highly natural synthesized speech for an arbitrary input character sequence.

1. In order to achieve the above object, the present invention reads the spectral parameters and residual signals registered in the speech segment file according to the phoneme sequence of the speech to be synthesized, and uses the read parameters and residual signals. In a speech synthesis method in which the spectral parameters are sequentially combined based on a certain rule and fed to a speech synthesizer to output speech, the spectral parameters are subjected to flattening processing,
The flattened parameters and the residual signal obtained from the inverse filter are registered in the speech segment file and synthesized. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is an electrical block diagram for explaining an embodiment of the present invention, and in the figure, 1 is an audio input section.

2 is an adaptive inverse filter, 3 is a spectrum analysis section, 4 is an inverse filter, and 5 is a speech segment file. In this embodiment, when creating a speech segment file, an adaptive inverse filter is passed through which the spectrum envelope is flattened. A residual waveform for generating a driving sound source signal is extracted by passing the audio waveform through an inverse filter of a synthesizer, and is registered together with the spectral parameters. This is what I did. First, the voice to be registered in the voice segment file 5 is input from the voice input unit 1,
After passing the audio waveform through an adaptive inverse filter 2, a spectrum analysis section 3 performs spectrum analysis. The adaptive inverse filter 2 is a filter having characteristics inverse to the vocal cord wave characteristics, and the spectral outline of the speech waveform that has passed through it is adaptively flattened.

It is already known that this filter can be realized by a combination of second-order and third-order inverse filters. In the spectrum analysis unit 3, parameters necessary for synthesis, such as <LSP (line spectrum pair), etc., are extracted for each frame period based on linear prediction analysis, and are registered as a parameter dictionary of the elemental piece.

Further, the residual signal is generated by applying the parameters determined by the above method to the inverse filter 4 of the synthesizer and passing the original input speech waveform through this filter. When registering the residual waveform, any processing may be performed so that necessary information at the time of synthesis can be easily retrieved.

FIG. 2 is a diagram illustrating an example of a speech synthesis device using the above-mentioned speech units F and A. In the figure, 11 is a character array input section;
2 is a syntactic analysis unit, 13 is a dictionary, 14 is a phoneme sequence generation unit,
15 is a parameter time series generation unit, 16 is a pitch pattern generation unit, 17 is a phoneme time series generation unit, 18 is a speech segment file, 19 is a synthesis filter, 20 is a D/A conversion unit, 21
is a speaker, and the input character sequence from the character string input section 11 is processed by the syntax analysis section 12 into reading 9 accent positions.

Analysis of intonation, etc. is performed, and the phoneme sequence generation unit 1
4 and the analysis result is passed to the pitch pattern generation section 16.

The phoneme sequence generation unit 14 generates a phoneme code sequence required for synthesis, and the parameter time sequence generation unit 15 generates
The parameters of the corresponding segment are read from the speech segment file 18 and are communicated smoothly. Based on intonation information etc., the pitch pattern generation unit 16 generates
The pitch period is determined, and the sound source time series generation unit 17 reads out the residual waveform of the corresponding segment from the speech segment file 18, processes the pitch period, etc., and generates the required driving sound source signal. The parameters and time series data of the sound source are input to the synthesis filter 19 and passed through the D/A converter 20.
The voice is output from the speaker 21 as a synthesized voice. As a result, spectral distortion at the segment connection portion is reduced, so that smooth audio output can be achieved.

FIG. 3 is an electrical block diagram for explaining another embodiment of the present invention. This embodiment includes a phoneme determining unit 6 that determines whether speech is voiced or unvoiced when creating a speech segment file. ,
A switch 7 that operates according to the judgment result of the phoneme judgment unit 6, and two types of filters 8 and 9 for high-frequency emphasis and low-frequency emphasis.
The following describes the case of two types of filter configurations that switch between voiced and unvoiced, but is not limited to this. First, the audio signal input at the audio input section 1 is The phoneme determining unit 6 determines whether the sound is voiced or unvoiced. voiced.

As a method for determining unvoicedness, a method using the slope of the spectrum envelope is known. The switch 7 is linked to the phoneme determining section 6 and selects between two types of filters 8 and 9. If it is determined that there is a voice, the filter? (
+6dB10ct) is selected, and the audio signal is subjected to high-frequency emphasis and then subjected to spectrum analysis. If it is determined that there is no voice, filter 8 (6dB10
ct ) is selected, bass-emphasized, and then spectrally analyzed. The operation of the spectrum analyzer 3 and the flow of subsequent processing are the same as in the embodiment shown in FIG.

FIG. 4 is an electrical block diagram for explaining another embodiment of the present invention. In this embodiment, a parameter correction section 10 is used when creating a speech segment file. , the speech to be registered in the speech segment file 5 is input from the speech human input section 1, and the spectrum analysis section 3 performs spectrum analysis. The spectrum analysis unit 3 outputs linear prediction coefficients and slope information of a spectrum envelope obtained from the coefficients. The parameter correction unit 10 receives the results of the spectrum analysis unit 3, corrects the linear prediction coefficients so as to remove the slope of the spectrum envelope, converts them into parameters required at the time of synthesis, and creates 7 speech segment files. Register for 5. Here, we have explained spectral analysis and parameter correction based on linear predictive analysis, but the invention is not limited to this. For example, it is also possible to use FFT (fast Fourier transform) in the spectral analysis section. Extraction of the difference waveform and subsequent processing can be performed in the same manner as in the embodiment shown in FIG.

Results As is clear from the above description, according to the present invention, since sound source characteristics are not included in the parameters of registered speech segments, highly natural synthesized speech can be obtained when synthesizing any sentence.

[Brief explanation of the drawing]

FIG. 1 is an electrical block diagram for explaining an embodiment of a speech synthesis method according to the present invention, and FIG. 2 is an electrical block diagram for explaining an embodiment of a speech synthesis device. 3 and 4 are electrical block diagrams for explaining other embodiments of the present invention, respectively. 1... Audio input section, 2... Adaptive inverse filter, 3...
- Spectrum analysis section, 4... Inverse filter, 5... Speech segment file, 6... Phoneme determination section, 7... Switch, 8, 9... Filter, 10... Parameter correction section , 11...Character array input section, 12...Syntax analysis section, 13...Dictionary. 14... Phoneme sequence generation unit, 15... Parameter time series generation unit, 16... Pitch pattern generation unit, 17...
・Phonological time series generation unit, 18...Speech series file, 1
9... Synthesis filter, 20... D/A converter, 21
...Speaker.

Claims (4)

[Claims] (1) The spectral parameters and residual signals registered in the speech segment file are read out according to the phoneme sequence of the speech to be synthesized, and the read parameters and residual signals are sequentially combined based on a certain rule. In the speech synthesis method, the spectrum parameters are flattened and the flattened parameters and the residual signal obtained from the inverse filter are registered in the speech segment file. , a speech synthesis method characterized by synthesis. (2) The speech synthesis method according to claim (1), wherein the spectral parameters are flattened using an adaptive inverse filter for removing radiation and sound source characteristics in the speech generation process. (3) The speech synthesis method according to claim (1), wherein the spectral parameters are flattened using two or more types of filters and a phoneme determining section for selecting them. (4) The speech synthesis method according to claim (1), characterized in that the spectral parameters are flattened by performing spectral conversion of the speech waveform and performing parameter correction to remove its slope.