JPH01152500A - Voice synthesizer - Google Patents

Abstract

PURPOSE: To continuously vary a pitch under simple pitch control by employing a means which interpolate a pulse train of one frame in a voiced sound section, and filling a representative period with 0's behind a pulse train when a pitch period is made long and making the pulse train overlap when the pitch period is made short. CONSTITUTION: In the voiced sound section, the means which interpolates the pulse train of one frame from e pulse train in an arbitrary representative pitch section in a frame; and the representative section is filled with 0's (no pulse) behind the pulse train when the pitch period is made long and the pulse train in the representative section is made to overlap when the pitch period is made short. Namely, when pitch control information 12 is longer than a pitch period 10, an interpolating circuit 7 performs interpolation while making the pitch period 10 long, but when short, the interpolating circuit 7 performs interpolation while making the pitch period 10 short. Consequently, discontinuance at the connection point of frames can be prevented.

Description

[Detailed description of the invention] [Field of industrial application] The present invention divides input audio into frames of a fixed time length, extracts pitches from the input audio, and generates multi-pulse signals for each pitch represented by a fixed number of driving sound source pulses. [Background Art] Conventionally, speech synthesis devices have been developed using speech waveform encoding methods and speech analysis and synthesis methods. There is. The former audio waveform encoding method uses the waveform itself, so it can obtain natural audio quality, but it is difficult to control the pitch1, and the pitch of the encoded original audio waveform is continuously variable. I was able to do it. Therefore, several stages of waveform data with different pitch periods were prepared and selected from among them. Such a method is used, for example, by Fushikida, Takashima, and Mitome: [A Study of Ruled Synthesis by Articulatory Segment Waveform Editing Method,” Proceedings of the Acoustical Society of Japan, September-October 1985, P179-180
) is explained in

In addition, the latter speech analysis and synthesis method extracts bit information from the original speech waveform by a linear prediction method, and the bit information can be continuously changed. However, the speech analysis and synthesis method uses a pseudo sound source as a sound source, and in the case of voiced sounds, it is driven by one pulse per pitch, and in the case of unvoiced sounds, it is driven by white noise. Such a method is described, for example, in Fumitada Itakura: "Speech synthesis method using line spectrum frequency as a parameter and its implementation into LSI" Nikkei Electronics 1981.2.2 P128-15.8).

[Problem that the invention seeks to solve]

The above-mentioned speech waveform encoding method in the conventional speech synthesizer uses the waveform itself, so natural speech quality can be obtained, but it is difficult to control the pitch, and the pitch of the encoded original speech waveform is difficult to control. could not be varied continuously. For this reason, several levels of waveform data with different pitch periods were prepared and selected from among them. However, in this case, there was a problem that the pitch was not continuously variable and that a large capacity memory was required to store audio waveform data in the memory on the pitch cycle side.

Furthermore, the speech analysis and synthesis method in the conventional speech synthesis apparatus extracts pitch information from the original speech waveform by a linear prediction method, and the pitch information can be continuously changed. However, in the speech analysis and synthesis method, a pseudo sound source is used as the sound source9, and voiced sounds are driven by one pulse per pitch, and unvoiced sounds are driven by white noise, so it is natural. There was a problem in that it was not possible to obtain a reasonable voice quality.

[Failure to solve the problem]

The speech synthesis device of the present invention is a speech synthesis device that uses a multi-pulse method that divides input speech into frames of a fixed time length, extracts pitches from the input speech, and represents each pitch with a fixed number of driving sound source pulses. , in a voiced section, an interpolation means is used to interpolate a pulse train for one frame from a pulse train in an arbitrary representative pitch section within the frame, and when the pitch period is lengthened in the voiced section, a zero is added after the pulse train in the representative pitch section. When the pitch period is shortened in the voiced section by means of overlapping the pulse trains, and when the pitch period is shortened in the voiced section, the means for overlapping the pulse trains is taken, and the means for correcting the change in voice power from frame to frame by the interpolation means. It has the following.

[Effect]

In the present invention, in a speech synthesis device using a speech waveform encoding method (multipulse method) having a speech generation model, in a voiced section, a pulse train for one frame is interpolated from a pulse train in an arbitrary representative pitch section within the frame. If the pitch period is to be lengthened, an O (no pulse) is placed behind the pulse train in the representative section, and if the pitch period is to be shortened, the pulse trains in the representative section are overlapped.

〔Example〕

Hereinafter, embodiments of the present invention will be described in detail based on the drawings.

Regarding the pitch interpolation type multi-pulse method using a representative pitch interval pulse train, for example, see Tezawa-Nori: "Real-time codec using a multi-pulse audio encoding method with little deterioration of sound quality even at bits per second or less" Nikkei. Electronics 1986.6°16 (A397) P185. −
215, the details will be omitted.

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

In the figure, reference numeral 1 is an audio file that stores an audio waveform obtained by analyzing an original audio signal using the pitch interpolation type multipulse method. FIG. 2 shows information stored in the audio file 1. Audio file 1 has a pitch period of 10 and P for each frame.
ARCOR coefficient 11 and sound source information 9 are stored. Here, in the case of a voiced section, the sound source information 9 is the amplitude and position information of the pulse train of the representative pitch section, and in the case of the unvoiced section, it is the amplitude and position information of the pulse train of the entire frame.

2 is a voiced/unvoiced judgment circuit; this voiced/unvoiced judgment circuit 2
If the pitch period 10 is set to O (zero), it is determined to be an unvoiced sound, and the sound source information 9 is sent to the pulse restoration circuit 3, and if the pitch period 10 is set to a value other than 0, it is determined to be a voiced sound. , sends the sound source information 9 to the bit control circuit 4.

In the case of an unvoiced section, the sound source information 9 is a pulse train of the entire frame, so it is input to the synthesis filter 6 as is. In the case of a voiced section, the sound source information 9 is a pulse train of the representative pitch section.

5 is an amplitude control circuit which receives the output of the pitch control circuit 4 as an input; 7 is an interpolation circuit which receives the pitch period 10 and pitch control information 12 and sends its output to the pitch control circuit 4; and 8 receives the PARCOR coefficient 11 as an input. An interpolation circuit 13 sends its output to the synthesis filter 6, and 13 is a synthesized sound that is the output of the synthesis filter 6.

The pulse restoring circuit 3, pitch control circuit 4, and amplitude control circuit 5 constitute a sound source pulse restoring section 14. The pitch control circuit 4 and the interpolation circuit T also include an interpolation means for interpolating a pulse train for one frame from a pulse train of an arbitrary representative pitch section in a frame in a voiced section, and a representative pitch when the pitch period is lengthened in the voiced section. The apparatus includes means for padding the end of the pulse train of the interval with zeros and means for overlapping the pulse trains of the representative pitch interval when the pitch period is shortened in the voiced interval. Further, the amplitude control circuit 5 constitutes means for correcting the change in audio power from frame to frame caused by the interpolation means.

FIG. 3 is an explanatory diagram showing pulse search and interpolation processing in a voiced section.

In the figure, 21 is two frames of the original audio waveform J,
The kth and (k+1)th frames are shown. The analysis conditions are that the sampling frequency is 8 KHz and the audio analysis frame length is 20 m5.

22 is a pulse train in a frame. This uses the pitch period determined for each frame to divide the frame into sub-frame sections for each pitch period, interpolates the coefficients of the synthesis filter determined for each frame for each sub-frame, and
Filter coefficients are determined for each frame, and a predetermined number of pulse trains are calculated for each sub-frame. The pulse train of each sub-frame of the k-th frame is (alkH12+'3kg 84, j15k
), the pulse train of each sub-frame of the (k+1)th frame is (k+1 k+l
h+11kl, a2+a3 e'4
+&S).

23 is a pulse train of a representative pitch section. This is selected by searching several sections for a section that can reproduce good audio that is very similar to the original audio throughout the frame. The representative pitch interval pulse train of the k-th frame is represented by (IL2), and the representative pitch interval pulse train of the (k+1)th frame is represented by (, +11).

Here, the pitch period 10 in the audio file 1 in FIG. 1 is the pitch period obtained by analyzing the original audio signal, and the pitch control information 12 is the value of the pitch period at which the pitch period 10 of the original audio signal is desired to be changed.

If the pitch control information 12 is longer than the pitch period 10, the interpolation circuit 7 performs interpolation while lengthening the pitch period 10, and if it is shorter, the interpolation circuit T performs interpolation while shortening the pitch period 10. The interpolation method will be explained below.

24 in FIG. 3 is a processed waveform showing a reproduction pulse train when the pitch period is lengthened, and FIG. 4 is an enlarged view thereof.

Pulse train I26 (
The shift period is extended by filling O after a2"1). After this, the pulse train lI27 (a2) of the (k+1)th representative pitch section is connected, and this is followed by the pulse train of the extended frame (bl +1 to b2, +1 to b3, +1 to b).As can be seen from this, no sound source pulse is set in the section where the pitch period is extended.

25 is a processed waveform showing a reproduction pulse train when the pitch period is shortened, and FIG. 5 is an enlarged view thereof.

Pulse train I26 (
The pitch period is compressed by overlapping a2"J and the pulse train U27 (at"") of the (k+1)th representative pitch section. After this, the pulse train U27 (at"") of the (k+1)th representative pitch section This is followed by the pulse train of the expanded frame (c
l+C! , c3 + to +l k+1. C
6) Generate. Then, in the C4+05 overlapping interval, both the pulse train I 26 of the representative pitch interval and the pulse train 1127 of the representative pitch interval may be set. This is because it is thought that voice quality will not deteriorate if the sound source information is not reduced by setting pulses even in the overlapped sections.

Here, the pulse train I26 in the representative pitch section and the pulse train [27] in the representative pitch section are the same pulse train.

The connection method at the boundary point between frames is such that the pitch section expanded or compressed in the k-th frame is (k
+1) Connect so that the expansion or compression of the th frame starts after the part that protrudes into the th frame. This prevents discontinuities at frame connection points.

In the voiced interval interpolation process described above, when the pitch period is lengthened, the pulse train in the frame decreases, so the sound source power becomes small, and when the pitch period is shortened, the pulse train in the frame increases, so the sound source power decreases. becomes larger. An example of an amplitude control method for averaging this sound source power will now be described.

When changing the pitch period, the amplitude control circuit 5 in FIG. 1 multiplies the amplitude of the pulse train by an amplitude correction value. Here, this amplitude correction value is a value obtained by dividing the number of pulses when the pitch period is not changed by the Hals number when the pitch period is changed.

The pulse train resulting from the amplitude correction performed in this manner is input to the synthesis filter 6. This synthesis filter 6 synthesizes audio using the PARCOR coefficients 11 of the audio file 1 which are interpolated for each subframe by the interpolation circuit 8. Here, we took the PARCOR coefficient as an example, but the α coefficient and LS
Other linear prediction coefficients such as the P coefficient may also be used.

〔Effect of the invention〕

As explained above, the present invention provides a speech synthesis device using a speech waveform encoding method (multipulse method) having a speech generation model, in which, in a voiced section, one frame's worth of pulses is extracted from a pulse train in an arbitrary representative pitch section within the frame. A simple pitch method that interpolates the pulse train, pads 0 (no pulse) after the pulse train in the representative section to lengthen the pitch period, and overlaps the pulse train in the representative section to shorten the pitch period. Depending on the control method, the pitch can be varied continuously.

Furthermore, this pitch control method can correct the instability of the audio amplitude caused by changes in the audio power for each frame.

Therefore, we used the multi-pulse method to convert the reference speech into shi phoneme segments and c.
If you prepare waveform data by breaking it down into synthesis units such as v-vc phoneme chains, and use the pitch control of this method during rule synthesis, the pitch will change smoothly and natural-looking rule synthesized speech can be obtained. There is an effect that it can be done.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing an embodiment of the present invention, Fig. 2 is an explanatory diagram showing information stored in the audio file shown in Fig. 1, and Fig. 3 is a pulse search and interpolation process in voiced sections. FIG. 4 is an enlarged view of the processing waveform when the pitch period is lengthened, and FIG. 5 is an enlarged view of the processing waveform when the pitch period is shortened. 1... Audio file, 2... Voiced/unvoiced determination circuit, 3... Pulse restoration circuit, 4... Pitch control circuit, 5... Amplitude control circuit, 6... - Synthesis filter, 7, 8... interpolation circuit. Patent applicant: NEC Corporation

Claims (1)

[Claims] In a speech synthesizer that uses a multi-pulse method, input speech is divided into frames of a fixed time length, pitches are extracted from the input speech, and each pitch is expressed using a fixed number of driving sound source pulses. An interpolation means is used to interpolate a pulse train for one frame from a pulse train in a representative pitch section, and when the pitch period is lengthened in the voiced section, a means is used to pad the pulse train in the representative pitch section with zeros, When the pitch period is shortened, the speech synthesis device takes means for overlapping the pulse trains of the representative pitch section, and has means for correcting the change in audio power from frame to frame by the interpolation means. .