JPH0315759B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech analysis and synthesis device that edits and synthesizes speech unit waveforms of approximately pitch periods extracted from natural speech.

Conventionally, the analysis unit takes advantage of the fact that the voice part of the speech waveform is a periodic waveform, and that the pitch periodic waveform (speech element waveform) changes relatively slowly in the vowel stationary part. A speech analysis and synthesis method is known in which a representative speech segment waveform is selected in a synthesis section, and a speech segment waveform is repeatedly used to generate a speech waveform in a synthesis section, thereby compressing the information amount of speech. The above method is described in, for example, literature (1).
Kuromatsu, Inoue; “Simulation of speech synthesis by recording and editing of pitch-unit speech segments” October 1970, Acoustical Society of Japan conference proceedings document (2-1-4) p.125-126
is described in detail.

However, the conventional method has the disadvantage that when the spectra of adjacent speech unit waveforms change rapidly, a large discontinuity occurs in the temporal change of the short-time spectrum of the synthesized waveform, resulting in a tendency for the synthesized sound quality to deteriorate. Furthermore, regarding the pitch period, if the difference between the time length of the extracted speech segment waveform and the pitch period when editing and synthesizing the speech segment waveform is large, the synthesized sound quality will deteriorate due to truncation of the segment waveform, etc. There are drawbacks.

An object of the present invention is to alleviate changes in sound quality caused by rapid changes in the spectral shape or pitch period between adjacent speech segment waveforms in a speech analysis and synthesis device that edits and synthesizes speech segment waveforms extracted from natural speech. The object of the present invention is to provide a speech analysis and synthesis device capable of producing synthesized speech of relatively high quality.

The present invention includes means for selecting the speech segment waveform as a representative speech segment waveform by comparing the spectral envelope parameters and pitch frequencies extracted from the natural speech waveform in pitch synchronization, that is, in synchronization with the pitch interval, in the analysis section. , means for generating a speech waveform between the two representative speech segments by weighting the preceding representative speech segment waveform and the following representative speech segment waveform and then adding them together and interpolating them in a synthesis unit; It consists of

The feature of the present invention is to select the speech segment waveform as a representative speech segment waveform by comparing the spectral envelope parameters and pitch frequencies extracted in pitch synchronization, and select the preceding representative speech segment waveform and the following representative speech segment waveform. The purpose of the present invention is to generate a speech waveform between the two representative speech segment waveforms by weighting the representative speech segment waveforms, then adding them together and interpolating them.

The representative speech segment waveform can be selected, for example, by pitch-synchronously extracting the spectral envelope parameter and pitch frequency from the natural speech and comparing these parameter values with the temporally preceding representative speech segment waveform. Let's do it. At this time, it is determined whether not only the distance between spectral envelope parameter values but also the distance between pitch frequencies exceeds a certain threshold, and if the distance between any of the parameter values exceeds the threshold, a new If the distance between any parameter values does not exceed the threshold value, it is considered to be the same as the preceding representative speech segment waveform, and it is not selected as the representative speech segment waveform, and a new speech segment waveform is selected as the representative speech segment waveform. The analysis is repeated until a representative segmental waveform is selected. At this time, a parameter representing the time interval over which the preceding representative speech segment waveform represents the speech segment waveform is extracted.

Next, a method for generating a composite waveform between representative speech units according to the present invention will be explained in detail.

The amplitude value of the temporally preceding representative speech unit waveform is f(t), and the amplitude value of the subsequent representative speech unit waveform is g
(t). (Here t represents time.) Weight α _i (here i=1,
2, . . . , n, which represents the number of repetitions of the pitch section. ), the subsequent representative speech segment waveform is multiplied by weight β _i (i=1, 2, . . . , n), and both are added to synthesize the waveform between the representative speech segments. That is, the synthesized speech segment waveform for one pitch is h _i
(t) (i=1, 2, . . . n), h _i (t) is given by the following equation.

h _i (t) = α _i · f (t) + β _i · g (t) ... (1) At this time, the weights α _i and β _i are changed as follows each time it is repeated.

１＞α ₁ ＞α ₂ ＞……＞α _o ＞0 O＜β ₁ ＜β ₂ ＜……＜β _o ＜1 …(2) Also, the pitch of the preceding representative speech segment waveform f(t) Since the period T _f and the pitch period T _g of the subsequent representative segment waveform g(t) generally have different time lengths, the pitch period T _i of the speech segment waveform h _i (t) for one pitch to be synthesized is is the pitch period obtained by interpolating the pitch periods T _f and T _g of the two representative speech unit waveforms. That is, for example, when T _f > T _g , T _f ≧T ₁ ≧...≧T _o ≧T _g ...(3), and when T _f < T _g , T _f ≦T ₁ ≦...≦T _o ≦T _g ……(4).

At this time, the pitch period of the representative speech segment waveform is
There may be cases where T _f or T _g is too long or too short with respect to the pitch period T _i of the speech unit waveform h _i (t) for one pitch to be synthesized. If the waveform is too short, the final amplitude value of the waveform is retained and used, and if the representative speech unit waveform is selected so that the difference in pitch period does not become large as in this method, the sound quality can be reduced to almost nothing. A composite waveform can be generated without deterioration.

In the above explanation, the pitch interval is repeated n times, but this is controlled according to the parameter representing the time interval represented by the preceding representative speech unit waveform extracted by the analysis section. In some cases, it may occur zero times, that is, in other words, a composite waveform between the two representative speech segment waveforms may not be generated.

As described above, according to the present invention, deterioration in synthesized sound quality due to sudden changes in the spectral envelope and pitch period between the two representative speech segment waveforms is alleviated,
It is clear that this method has the effect of obtaining a relatively high quality synthesized sound.

Next, embodiments of the present invention will be described using the drawings.

The figure is a block diagram showing one embodiment of the present invention.

First, natural speech is input to the speech waveform temporary storage circuit 10 via the natural speech input terminal 103 of the analysis section 101.
4. Input to pitch extraction circuit 105 and spectrum envelope information extraction circuit 106. Regarding the pitch extraction circuit, see, for example, the literature [2] LRRabiner et al.
Comparative Performance of Several Pitches
Detection Algorithms” IEEE Trans.Assp−24
Some methods and block diagrams are described in detail on No. 5, pages 399-418. Also, regarding the spectral envelope information extraction circuit, see, for example, document [3] LR
Rabiner et al. “Digital Processing of Speech
Since several systems, block diagrams, etc. are described in detail in Chapters 6 to 8 of "Signals" Prentice-Hall 1978, detailed explanations will be omitted.

The pitch information extracted by the pitch extraction circuit 105 in accordance with the control information sent from the analyzer control circuit 120 via the pitch extraction circuit control information transmission line 145 is sent to the pitch information comparison circuit 110 and the representative via the pitch information transmission line 108. It is sent to the speech segment waveform extraction circuit 121.

Spectrum envelope information extraction circuit 1 according to control information sent from analysis unit control circuit 120 via spectrum envelope information extraction circuit control information transmission line 146
The spectrum envelope information extracted in step 06 is sent to the spectrum envelope information comparison circuit 112 via the spectrum envelope information transmission line 109.

In the pitch information comparison circuit 110, the analysis section control circuit 120 is connected to the pitch information comparison circuit control information transmission line 1.
16, the pitch information sent from the pitch extraction circuit 105 and the pitch information of the temporally preceding representative speech unit waveform stored in the pitch information storage circuit 111. The comparison result information as to whether the absolute value of the difference exceeds a certain reference value is sent to the determination circuit 123 via the pitch information comparison information transmission line 115.
send to

In the spectral envelope information comparison circuit 112, according to the control information sent from the analysis section control circuit 120 via the spectral envelope information comparison circuit control information transmission line 119, the spectral envelope information sent from the spectral envelope information extraction circuit 106, Compare the spectral envelope information of the temporally preceding representative speech unit waveform stored in the spectral envelope information storage circuit 113 to determine whether the distance between the two spectral envelope information exceeds a certain reference value. Comparing information to spectrum envelope information Information transmission path 11
8 to the determination circuit 123.

In the determination circuit 123, the analysis section control circuit 12
0 through the determination circuit control information transmission path 124, when at least one of the comparison result information sent from the two comparison circuits indicates that it exceeds the reference value. It is determined that the input speech waveform is to be adopted as a new representative speech segment waveform, and if the information of the two comparison results indicates that both do not exceed the reference value, the new representative speech segment waveform is adopted. It is determined that the information is not to be adopted as
14, the pitch information comparison circuit 110, the spectrum envelope information comparison circuit 112, the representative audio waveform extraction circuit 121, and the counting circuit 148 via the judgment information transmission line B117, the judgment information transmission line C122, and the judgment information transmission line D147, respectively. send to

In the pitch information comparison circuit 110 and the spectrum envelope information comparison circuit 112, the determination circuit 123
When judgment information indicating that the waveform has been adopted as a new representative speech unit waveform is sent, the pitch information and spectral envelope information of the new representative speech unit waveform are stored in the pitch information storage circuit 111 and the spectral envelope information storage circuit, respectively. Write to 113.

The representative speech segment waveform extraction circuit 121 extracts the pitch according to the control information sent from the analysis section control circuit 120 via the representative speech segment waveform extraction circuit control information transmission line 125 and the judgment information sent from the judgment circuit 123. Based on the pitch information sent from the circuit 105, a representative speech segment waveform is cut out from the input speech waveform sent from the input speech waveform temporary storage circuit 104 via the input speech waveform transmission path 107, and the representative speech segment waveform data is analyzed. The pitch information of the representative speech segment waveform is outputted to the analysis section representative speech segment waveform pitch information output terminal 127.

The counting circuit 148 determines whether an adjacent new The number of determination information indicating that the waveform is not adopted as a representative voice unit waveform that occurs between the determination information indicating that the waveform has been adopted as a representative voice unit waveform is counted, and the number of determination information indicating that the waveform is not adopted as a representative voice unit waveform is calculated. It is output to the analysis section representative speech unit waveform repetition information output terminal 128 as repetition information indicating whether the speech waveform is representative.

The analysis unit control circuit 120 includes the pitch information comparison circuit 110, the spectrum envelope information comparison circuit 112,
The determination circuit 123 and the representative speech segment waveform extraction circuit 121 are controlled.

In the synthesis section 102, the representative speech segment waveform data is input to the synthesis section representative speech segment waveform data input terminal 12.
9 and stored in the representative speech segment waveform data storage circuit 132, and the representative speech segment waveform pitch information is inputted to the representative speech segment waveform pitch information input terminal 130 of the synthesis section and stored in the representative speech segment waveform pitch information storage circuit. The representative speech unit waveform repetition information is inputted to the representative speech unit waveform repetition information input terminal 131 of the synthesis section and stored in the representative speech unit waveform repetition information storage circuit 134.

The weight and pitch period calculation circuit 138 calculates the representative voice from the representative speech segment waveform pitch information storage circuit 133 in accordance with control information sent from the synthesis unit control circuit 140 via the weight and pitch period calculation circuit control information transmission line 139. The pitch information of the preceding representative speech segment waveform and the pitch information of the subsequent representative speech segment waveform are obtained via the segment waveform pitch information transmission path 136, and the pitch information of the representative speech segment waveform that follows is obtained from the representative speech segment waveform repetition information storage circuit 134. Repetition information of the preceding representative speech element waveform is obtained via the speech element wave repetition information transmission path 137, and the weight value and pitch period for each repetition section are calculated by equations (2) and (3) or ( 4) is calculated and sent to the amplitude and pitch period adjustment circuit 152 via the weight value transmission path 150 and the pitch period transmission path 151, respectively.

The amplitude and pitch period adjustment circuit 152 selects the representative speech segment from the representative speech segment waveform data storage circuit 132 in accordance with control information sent from the synthesis section control circuit 140 via the amplitude and pitch period adjustment circuit control information transmission line 153. Waveform data transmission line 1
35, obtain the preceding representative speech segment waveform data and the subsequent representative speech segment waveform data, multiply each speech segment waveform data by the weight sent from the weight and pitch period calculation circuit 138, and calculate the weight. If the pitch period of each representative speech unit waveform data is short with respect to the pitch period sent from the pitch period calculation circuit 138, the final amplitude value is held; if it is long, the pitch period is adjusted by cutting off the excess waveform. Each weighted speech unit waveform data is sent to an adding circuit 155 via a speech unit waveform data transmission line 154.

The adder circuit 155 combines the weighted preceding speech segment waveform sent from the amplitude and pitch period adjustment circuit in accordance with control information sent from the synthesizer control circuit 140 via the adder circuit control information transmission line 156. Generate an interpolated waveform by adding the amplitude values at the corresponding times of the subsequent speech segment waveform data,
The editing synthesis circuit 14 via the interpolation waveform transmission line 141
Send to 2.

The editing/synthesizing circuit 142 selects the representative speech segment waveform data transmission path 135 from the representative speech segment waveform data storage circuit 135 in accordance with the control information sent from the synthesis unit control circuit 140 via the editing/synthesizing circuit control information transmission path 143. editing and synthesizing the representative speech unit waveform data sent through the representative speech unit waveform data and the interpolation waveform sent from the representative speech unit waveform interpolation circuit 138;
It is output to the synthesized speech output terminal 144.

The synthesis section control circuit 140 controls the weight and pitch period calculation circuit 138, the amplitude and pitch period adjustment circuit 152, the addition circuit 155, and the editing and synthesis circuit 142.

In the embodiment shown in the figure, the representative speech segment waveform data, representative speech segment waveform pitch information, and representative speech segment waveform repetition information obtained by the analysis section are stored in the storage circuit. It is also possible to realize an audio communication device by synthesizing in real time without storing it in a memory circuit.

[Brief explanation of the drawing]

The figure is a block diagram showing one embodiment of the present invention. In the figure, 101 is an analysis section, 102 is a synthesis section, 103 is a natural speech input terminal, 104 is an input speech waveform temporary storage circuit, 105 is a pitch extraction circuit,
106 is a spectral envelope information extraction circuit, 110 is a pitch information comparison circuit, 111 is a pitch information storage circuit, 112 is a spectral envelope information comparison circuit, 11
3 is a spectrum envelope information storage circuit, 120 is an analysis section control circuit, 121 is a representative speech unit waveform extraction circuit, 123 is a determination circuit, 1126 is an analysis section representative speech unit waveform data output terminal, and 127 is an analysis section representative phoneme. Single waveform pitch information output terminal, 128 is a representative voice unit waveform repetition information output terminal of the analysis section, 12
9 is a representative voice unit waveform data input terminal of the synthesis section; 1
30 is a representative speech unit waveform pitch information input terminal for the synthesis section, 131 is a representative speech unit waveform repetition information input terminal for the synthesis section, 132 is a representative speech unit waveform data storage circuit, and 133 is a representative speech unit waveform pitch information storage circuit. , 134 is a representative speech unit waveform repetition information storage circuit, 138 is a weight and period calculation circuit, 14
0 is a synthesis unit control circuit, 142 is an editing synthesis circuit, 1
44 is a synthesized voice output terminal, 148 is a counting circuit, 1
Reference numeral 52 represents an amplitude and pitch period adjustment circuit, and 155 represents an addition circuit.

Claims (1)

[Claims] 1. In a speech analysis and synthesis device of the type that edits and synthesizes a speech segment waveform with a pitch period or so extracted from a natural speech signal, the spectral envelope parameter value and pitch frequency corresponding to the extracted speech segment waveform are predetermined. means for selecting a new spectral envelope parameter, a pitch frequency, and a representative speech segment waveform from among the speech segment waveforms by comparing with respective values corresponding to the representative speech segment waveform;
It is characterized by having means for generating a speech waveform between the two representative speech unit waveforms by weighting and adding together the preceding representative speech unit waveform and the following representative speech unit waveform. Speech analysis and synthesis equipment.