JP2709198B2 - Voice synthesis method - Google Patents

Abstract

PURPOSE:To realize a smaller capacity storage device, control the speed of uttered voice and pitch of the voice and to obtain a high quality synthesized voice in a waveform editing/synthesizing method. CONSTITUTION:Each waveform P1, Pm and Pn, which includes three high peaks of original sound phoneme waveform column P1 to Pn+1 shown in A, is stored in a storage device as a representative waveform, similarly representative waveforms of each phoneme are stored, input synthesized text is analyzed to obtain phoneme columns, the representative waveforms of each of the phonemes are read out form the storage device and output synthesized waveforms P1m(0), P1m(1),..., P1m(k) which are obtained by interpolating the read out representative waveforms P1, Pm and Pn and their respective adjacent waveforms P1 and Pm employing given synthesized pitch period Tp and connecting time L, where P1m(i) (i=1, 2,...k) is P1Xalpha(i)+PmXbeta(i), alpha(i) and beta(i) are weighting coefficients against the waveforms P1 and Pm, respectively and alpha(i)=0.5X[1+cos{piX(L-Ti)/L}] and beta(i)=1-alpha(i). Ti is the time interval between P1m(0) and P1m(i).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is applied to a method of synthesizing an arbitrary speech word according to a rule, a method of compressing and transmitting information of a speech signal, transmitting or storing the compressed speech signal, and resynthesizing the compressed speech signal. The present invention relates to a speech synthesis method for synthesizing speech from waveform information.

[0002]

2. Description of the Related Art A large-scale storage device is required to store all waveform information of a voice in a storage device and implement a voice synthesis method using the storage device to perform voice synthesis.
The method of synthesizing arbitrary spoken words reduces storage capacity,
Conventionally, in order to output a variety of synthesized sounds, only the feature parameters of the voice extracted from the voice waveform are stored in a storage device, and the feature parameters are read out as needed to create a synthesized voice, so-called parameter editing synthesis. The method was mainly used. Also, a waveform editing / synthesis method for storing waveform information as it is has been used for the purpose of obtaining a high-quality synthesized sound.

[0003] However, in the above parameter editing / synthesis method, although various parameter controls are easy, the quality of the synthesized sound obtained is largely different from that of the real voice. In the waveform editing / synthesizing method, a high-quality synthesized sound can be obtained, but a large-capacity storage device is required (for example, Japanese Patent Application No. 63-115721). There is an example (Japanese Patent Application No. 56-179915) applied to a residual signal obtained by linear prediction analysis as a method using only a representative waveform of a voice section instead of storing all voice waveforms. ,
This does not use a waveform interpolation method that is actively synchronized with the pitch. An example in which the waveform interpolation method synchronized with pitch is used for connection between segments of CV (consonant-vowel) / VC (vowel-consonant) (Mitsuru et al., May 1981, lecture by The Acoustical Society of Japan) Papers, p431), but there is no synthesis method for information compression and high quality for the entire synthesis unit.

An object of the present invention is to provide a method of synthesizing a waveform to obtain a speech word, in which the storage capacity can be significantly reduced, a high-quality synthesized sound can be obtained, and the speech can be obtained with high efficiency. It is an object of the present invention to provide a speech synthesizing method capable of obtaining a high-quality synthesized sound from a compressed and transmitted and stored sound.

[0005]

According to the present invention, a plurality of representative waveforms extracted from different points in time of a speech waveform are obtained.
According to a given pitch period and duration, a voice waveform between the representative waveforms is interpolated using these representative waveforms to resynthesize a continuous voice. As described above, in order to obtain a high-quality synthesized sound by the waveform editing type speech synthesis method, it is necessary to previously store a speech waveform required for synthesis in a storage device. In comparison with the above, the scale of the storage device becomes very large, and an efficient measure for reducing the waveform information is required to realize economy and downsizing. Further, in the related art, it is very difficult to control the pitch period and the duration without deteriorating the quality of the synthesized voice, that is, to control the level of the synthesized voice and the utterance speed. In the present invention, this audio waveform is efficiently compressed by a method using only the selected representative waveform, paying attention to the pitch period of the audio, and is processed and reproduced so that distortion is not detected audibly during synthesis. In addition to the effect of reducing the storage capacity, it is possible to synthesize a voice having a changed pitch period or duration with high quality.

[0006]

FIG. 1 shows an embodiment in which the present invention is applied to a waveform editing / synthesizing method. This speech synthesis has two stages: an analysis and accumulation process and a synthesis process. First, in the voice analysis process, the analysis unit 11 inputs the original voice signal from the voice input terminal 12 and the phoneme labeling information from the labeling information input terminal 13 to the optimum waveform peak position search circuit 14.
The input original audio signal is segmented for each phoneme by inspection, and phoneme symbols and position information are added.

Next, a near-peak waveform extracting circuit 15
For each phoneme, about 3 representative waveforms
Search for the peak position and cut off the waveform centered on this position information.
And use this as a representative waveform for its position information and phoneme information.
And the like, and accumulate them in the storage device 16. For example, as shown in FIG. 2A
The original speech waveform sequence P that forms one phoneme₁P_TwoP _Three
… P_m… Are input, and from these are different as shown in FIG. 2B.
Three representative waveforms P₁, P_m, P_nIs extracted
You.

In the speech synthesis process, the synthesizer 21
A synthesized text is input from a text input terminal 22 to a text analysis unit 23, the synthesized text is analyzed and converted into a phoneme sequence, and a waveform readout circuit 24 reads a necessary speech waveform from the phoneme sequence from the storage device 16 and generates a pitch. It is supplied to the synchronous interpolation circuit 25. The pitch synchronous interpolation circuit 25 interpolates the speech waveform between the representative waveforms in synchronism with the synthesized pitch cycle and superimposes the speech waveforms on a pitch-by-pitch basis in accordance with the synthesized pitch cycle and the connection time provided from the prosody information generating circuit 26. obtain. At this time, the overlapping window shape for each pitch is
You can use trapezoidal windows and cosine function windows.

For example, a representative waveform P shown in FIG.₁, P_mBut
Readout composite pitch period T_P, Given the duration L
, The representative waveform P₁And P_mL is the time between
Period T _PIs interpolated from both waveforms as
P_1m(0), P_1m(1), P_1m(2)… P_1m(k)
obtain. P_1m(i) = P₁× α (i) + P_m× β (i) (1) where P_1m(0) = P₁, P_1m(k) = P_mAnd α
(I) and β (i) are representative waveforms P₁And P_mHeavy against
And expressed by equations (2) and (3), respectively.

Α (i) = 0.5 × [1 + cos {π × (L−Ti) / L}] (2) β (i) = 1−α (i) (3) where i is a composite waveform Number and Ti are P _lm (0) to P
time interval _lm (i), L represents a time interval in a representative waveform _{P l} and _{P m.} Similar interpolation processing is temporally adjacent
The following representative waveforms _Pm and _Pn are performed. In this way, a synthesized waveform obtained by interpolating between two representative waveforms that are temporally adjacent to each other is output as a continuous sound by the connection and synthesis circuit 27 for each phoneme.

In the above processing, the representative waveform selection method includes, for example, dividing the phoneme section into a plurality of sub-sections,
A method of obtaining a waveform peak for each of the sub-intervals and using the waveform as a representative waveform, or a method of selecting the spectrum information or the dynamic characteristics of the waveform information of the relevant phoneme interval as a guide, that is, a method of selecting the largest peak in the phoneme interval A method is conceivable in which the difference between adjacent waveforms is sequentially determined on the front side and the rear side, and the waveform of the portion that greatly changes and the maximum peak waveform are used as the representative waveform. In addition, a simple linear function such as Expression (4) can be used as the weight coefficient α (i).

Α (i) = (L−Ti) / L (4) In this embodiment, a speech synthesis method using phonemes as a unit has been described as an example, but it is also used in speech synthesis based on other synthesis units such as syllables. Clearly, it is possible. Further, the present invention can be applied to speech synthesis in high-efficiency speech coding. In other words, only the representative waveform in FIG. 2B is encoded and transmitted or stored, the information is compressed in the transmission path or the storage device, and the synthesized voice is analyzed on the receiving side or the storage and reading side according to the above-described method of the present invention. Is obtained. Furthermore, in the case of waveform editing speech synthesis, several similar representative waveforms are further represented by one waveform by using a clustering technique such as a vector quantization method for the entire representative waveform obtained for all synthesis units. Also, the information compression ratio can be increased.

In the above description, one phoneme has three representative waves.
Although the shape is used, it is sufficient if there are two or more representative waveforms.

[0014]

As described above, according to the present invention, since the voice waveform information can be efficiently compressed and reproduced by using the representative waveform, the storage capacity can be greatly reduced in the waveform editing type voice synthesis system and the transmission can be performed. / Accumulation enables high-efficiency encoding and obtains a high-quality synthesized sound. Further, by performing the waveform interpolation processing in synchronization with the pitch cycle and selecting the connection time, it is possible to control the utterance speed and the level of the voice without deteriorating the quality of the original voice.

[Brief description of the drawings]

FIG. 1 is a configuration diagram showing one embodiment of the method of the present invention.

FIG. 2 is a diagram showing examples of an original waveform, a representative waveform, and a composite waveform.

Claims (1)

(57) [Claims] 1. A is stored in the storage device a speech waveform of each speech unit, the synthesis, a method for synthesizing a continuous speech voice corresponding to the speech unit sequence given waveform read from the storage device, The storage device stores a pitch from the audio waveform in the audio unit.
From the waveform obtained corresponding to the
Select and store at least two or more waveforms as representative waveforms
And, at the time of synthesis, corresponding to the speech unit of the speech unit sequence given
The representative waveform is read from the storage device , and a set of two temporally adjacent representative waveforms in the representative waveform is
Speech waveforms between two representative waveforms were given for each
Given to be continuous for a length corresponding to the duration
A sequence of waveforms synchronized with the pitch period is converted to the two representative waveforms.
Generated by weighting synthesis
By interpolating between two representative waveforms adjacent to each other,
A speech synthesis method comprising synthesizing a speech waveform for each voice.