JPS5948396B2 - Fragment editing type speech synthesis device - Google Patents

Abstract

PURPOSE:To avoid a big discontinuity occurring to the frequency spectrum shape of the compounded voice and thus to improve the sound quality, by using the interpolation element piece waveform and the element piece waveform extracted from the natural voice waveform.

Description

[Detailed description of the invention] The present invention relates to a segment editing type speech synthesis device.

A type of speech waveform that extracts in advance various speech waveforms with a time length comparable to a pitch interval from a natural speech waveform, and edits and synthesizes the segment waveforms according to control data such as segment number and pitch prepared in advance as synthesis data. A segment editing type speech synthesis method is known. In addition, in the above method, when using segment waveforms extracted from limited natural speech waveforms such as monosyllables, the shape of the frequency spectrum of the segment waveform series used during synthesis changes sufficiently smoothly. There may be cases where the waveform does not contain enough elemental waveforms to obtain the desired result. This causes a large discontinuity in the frequency spectrum shape of the synthesized speech, which causes deterioration in sound quality. Furthermore, extracting an elemental waveform having a frequency spectrum shape necessary to compensate for the discontinuity of the frequency spectrum from a natural speech waveform is as follows:
The disadvantage is that it is extremely difficult to select natural speech having the required frequency spectral shape. The purpose of the present invention is to improve sound quality by preventing large discontinuities from occurring in the frequency spectrum shape such as the formant frequency of speech synthesized using segmental waveforms extracted from natural speech waveforms such as monosyllables and words. An object of the present invention is to provide a segment editing type speech synthesis device.

The speech synthesis device of the present invention includes means for calculating a parameter value representing the frequency spectrum shape of a speech segment waveform extracted from natural speech, and a parameter representing the frequency spectrum shape calculated from two different speech segment waveforms. A speech segment waveform extracted from the natural speech and a speech segment waveform generated by a speech segment interpolation device including means for interpolating values and a means for synthesizing a speech segment waveform from parameter values representing the interpolated frequency spectrum shape. and a means for editing and synthesizing the speech waveform using the speech segment waveforms obtained in conjunction with the speech segment waveforms.

A feature of the present invention is that in addition to the segment waveform extracted from a natural speech waveform such as a monosyllable, parameter values representing the frequency spectrum shape such as the formant of the segment waveform extracted from the natural speech waveform are interpolated. The purpose is to also use a segment waveform (herein referred to as an interpolated segment) generated from the values.

Therefore, when the present invention is used, large discontinuities do not occur in the spectrum of formants, etc. of synthesized speech, and synthesized speech with relatively good sound quality can be obtained. It is known that the frequency spectrum characteristic of a voice waveform includes a frequency component called a formant, which has greater energy than surrounding frequency components. Furthermore, it is known that when the formant frequency suddenly changes greatly in a vowel stationary region, the sound quality deteriorates. Therefore, as shown in FIG. 1, the interpolated segment waveform is an interpolated segment waveform 105 extracted from the natural speech to be interpolated (a single tone is shown in the figure) and an interpolated segment 106 that smoothly connects the IOT formant frequency. is suitable. In the present invention, the formant frequency and band width are extracted from the two end segment waveforms to be interpolated, and the interpolated segment waveform is generated from the value obtained by (linearly) interpolating the extracted formant frequency and band width. do. Also known is a method of extracting formant frequencies and bandwidths using linear prediction coefficients obtained by solving simultaneous linear equations having autocorrelation values of speech unit waveforms as coefficients. Conversely, a method of calculating a linear prediction coefficient from formant frequency and band width is also known, and the calculation method is described in detail in the following literature (magazine), so the explanation will be omitted here. Literature ``TheJOurnalOf
theAcAusticalSOcietyOfAme
rica'', VOl.5O, Part 2, Part 6
Pages 37-655 (titled “SpeechAnalys”)
isandSynthesisbyLinearPre
dichtiOnOftheSpEechWa-Ve''
, 1971TheJOwnalOftheAc10
ustica1S0city0fAmerica) The pair of formant frequency (f) and band width (b) extracted from the speech unit waveform at both ends (A, B) are (FAn, bAn), n = 1, 2, ・
, N and ( FBn, bBn) , n = 1, 2,
・ ・ ,N.

It is known that it is sufficient if N is about 4 to 6. The pair of formant frequency and band width obtained by (linear) interpolation of the formant frequency and band width at both ends is (FOn, bcn)n = 1, 2,...
. , N, for example, (FOn, fcn) can be given as shown in the following equation. By appropriately giving 1 and m in equation (1) and calculating Fcn and bOn, the formant frequency and band width of the interpolation element can be obtained.

The interpolated segment waveform is given as an impulse response waveform of a recursive filter controlled by linear prediction coefficients calculated from Fcn and bcn. In addition to the above-mentioned formant frequency and band width, parameters representing the frequency spectrum shape include the autocorrelation coefficient or linear prediction calculated when extracting the formant frequency (or polar frequency) and band width from the natural speech waveform. It is clear that coefficients can also be used.

However, in the method of using an autocorrelation coefficient or a linear prediction coefficient, the formant frequency of the interpolated segment does not necessarily result in a smooth interpolation of the formant frequencies of the segments extracted from the natural speech at both ends. In addition, the first
In the figure, reference numeral 101 is the time axis, reference numeral 102
is a frequency axis, reference numeral 103 indicates a change curve of the first formant, and reference numeral 104 indicates a change curve of the second formant, respectively.

Next, the present invention will be explained in detail with reference to the drawings.

FIG. 2 is a block diagram showing one embodiment of the present invention. First, a speech waveform such as a monosyllable is inputted from the speech waveform input terminal 201 to the speech segment extraction section 206, and is processed according to control data given from the leg control circuit 202 via the speech segment extraction section control data transmission line 203. A speech segment waveform is generated.

The speech segment waveform generated in the speech segment extraction section 206 is output to the interpolation segment generation section 201, and at the same time, the speech segment waveform is outputted to the segment waveform storage circuit 214 in the synthesis section 210 by the switch 20.
8. In the interpolation segment generation section 20T, a control data transmission line 204 is transmitted from the control circuit 202 to the interpolation segment generation section control data transmission line 204.
The interpolated segment waveform is calculated according to the interpolated segment generation data given via the segment waveform storage circuit 2 in the synthesis unit 210.
14 via the switch 209. After the above operations are completed, the audio waveform is synthesized in accordance with the output command data provided from the control circuit 202 to the synthesis unit 210 via the output command data transmission line 205. A synthesis unit control circuit 211 in the synthesis unit 210 connects the output command data transmission path 21
The synthesized data storage circuit control data and the editing synthesis circuit control data are generated in accordance with the output command data given via the synthesis data storage circuit control data transmission line 212 and the editing synthesis circuit control data transmission line 213, respectively. Synthesis data storage circuit 215 and editing and synthesis circuit 21
Control 6. The composite data storage circuit 215 outputs the segment number data stored in advance to the segment waveform storage circuit 214 according to the composite data storage circuit control data, and at the same time edits and synthesizes the pitch and amplitude data stored in advance. Output to. The segment waveform storage circuit 214 stores a segment consisting of a speech segment waveform extracted directly from the natural speech waveform according to the segment number data and provided via a switch 208 and an interpolated segment waveform provided via a switch 209. The waveform is output to the editing/synthesizing circuit 216. The editing/synthesizing circuit 216 outputs data from the segment waveform storage circuit 214 in accordance with the editing/synthesizing circuit control data given via the editing/synthesizing circuit control data transmission line 213 and the pitch and amplitude data output from the synthetic data storage circuit 215. A synthesized waveform is generated by editing and synthesizing the segment waveforms, and the synthesized waveform is output from the synthesized waveform output terminal 217. Next, a method for generating an interpolated segment waveform will be explained with reference to FIG.

FIG. 3 is a block diagram showing in detail the interpolated segment waveform generating section 207 of FIG. First, a segment waveform extracted from two natural voices to be interpolated is inputted from the segment waveform input terminal 306 to the formant extraction circuit 307 .

The formant extraction circuit 307 extracts the formant frequency and bandwidth from the elemental waveform extracted from the natural speech according to control data given from the control circuit 301 via the formant extraction circuit control data transmission line 302, and extracts the formant frequency and band width from the segment waveform extracted from the natural speech. Output to. The interpolation circuit 308 interpolates the formant frequency and band width output from the formant extraction circuit 307 according to control data given from the control circuit 301 via the interpolation circuit control data transmission path 303, and calculates the interpolated value of the formant frequency (( 1)
Fcn) in the formula and the interpolated value of the band width ((1)
Bcn) in the formula is calculated, and the predictive side coefficient calculation circuit 309
Output to. The prediction coefficient calculation circuit 309 is the control circuit 301
A prediction coefficient calculation circuit calculates a linear prediction coefficient from the formant frequency interpolation value and band width interpolation value output from the interpolation circuit 308 according to control data given via the control data transmission line 304, and
Output to 0. On the other hand, the sound source waveform generation circuit 311 is connected from the control circuit 301 to the sound source waveform generation circuit control data transmission line 305.
An impulse waveform is generated as a sound source waveform based on control data given via the control data, and is output to the linear filter 310. The linear filter 310 is controlled by the linear prediction coefficient outputted from the prediction coefficient calculation circuit 309, calculates an interpolated segment waveform using the impulse waveform outputted from the sound source waveform generation circuit 311 as a sound source, and outputs an interpolated segment waveform to an interpolated segment waveform output terminal. 3
12 outputs an interpolated segment waveform.

[Brief explanation of the drawing]

FIG. 1 is a diagram for explaining the principle of the present invention, FIG. 2 is a block diagram showing an embodiment of the present invention, and FIG. 3 is a block diagram showing in detail the interpolated segment waveform generating section 207 in FIG. In Fig. 2, 201 is the audio waveform input terminal;
02 is a control circuit, 203 is a speech segment extraction unit control data transmission line, 204 is an interpolation segment generation unit control data transmission line, 20
5 is an output command data transmission path, 206 is a speech segment extraction unit,
207 is an interpolation segment generation unit, 208 is a switch, 209 is a switch, 210 is a synthesis unit, 211 is a synthesis unit control circuit,
212 is a synthetic data storage circuit control data transmission line; 213
214 is a segmental waveform storage circuit, 215 is a composite data storage circuit, 216 is an editing and composite circuit, and 217 is a composite waveform output terminal.

Claims (1)

[Claims] 1. In a segment editing type speech synthesis device that edits and synthesizes phoneme forms of a plurality of pitch intervals extracted from a natural speech waveform in advance, a parameter representing the frequency spectrum shape of a speech segment waveform extracted from the natural speech. means for calculating a value, means for interpolating a parameter value representing the frequency spectrum shape calculated from two different speech segment waveforms, and synthesizing a speech segment waveform from the parameter values representing the interpolated frequency spectrum shape. 1. A speech segment editing type speech synthesis device characterized in that a speech segment waveform generated by a speech segment interpolation device including means for performing the above-mentioned natural speech is used in conjunction with a speech segment waveform extracted from the natural speech.