JP2937322B2 - Speech synthesizer - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a speech synthesizer that performs pitch control.

[Prior art] Conventionally, the speech waveform encoding method uses a waveform itself, so that natural speech quality can be obtained, but it is difficult to control the pitch. Could not be changed. Therefore, several steps (for example, five steps) of waveform data having different pitch periods are prepared and selected from among them (for example, Reference 1;
Fushida, Takashima, Midome: "A Study on Rule Synthesis by Articulatory Segment Waveform Editing Method", Proc. Of the Acoustical Society of Japan, September-October pp. 179-180).

In the speech analysis / synthesis method, pitch information is extracted from an original speech waveform by a linear prediction method, and the pitch information can be continuously changed. However, in the voice analysis / synthesis method, a pseudo sound source is used as a sound source.
It was driven by one pulse per pitch, and in the case of unvoiced sound, it was driven by white noise. (For example, Ref. 2; Fumitada Itakura: "Speech synthesis method using line spectrum frequency as a parameter and its LSI implementation" Nikkei Electronics 1981.2.2P128
-158).

[Problems to be Solved by the Invention] However, since the speech waveform encoding method uses the waveform itself, natural speech quality can be obtained. However, it is difficult to control the pitch, and the pitch of the encoded original speech waveform cannot be changed continuously. Therefore, several stages (five stages) of waveform data having different pitch periods are prepared and stored, and selected from among them. However, in this case, there is a disadvantage that a large-capacity memory is required because the voice waveform data is stored in the memory for each pitch cycle instead of continuously changing the pitch.

In the speech analysis / synthesis method, pitch information is extracted from an 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 onset, and in the case of voiced sound, one pulse is driven at one pitch, and in the case of unvoiced sound, white noise is used. However, there is a disadvantage that natural voice quality cannot be obtained.

SUMMARY OF THE INVENTION An object of the present invention is to provide a speech synthesizing apparatus capable of eliminating the above-mentioned drawbacks, continuously changing the pitch, and obtaining a natural speech quality.

[Means for Solving the Problems] According to the present invention, a pitch is extracted from the input voice in a state where the input voice is divided into frames of a fixed time length, and is represented by a certain number of drive sound source pulses for each pitch. In a speech synthesizer using a multi-pulse method, a representative pitch section in a frame of two types of speech waveforms having different pitch frequencies (a speech waveform having a high pitch frequency and a speech waveform having a low pitch frequency) analyzed by the multi-pulse method. Linear interpolation of the amplitude information of the pulse train, the position information, and the linear prediction coefficient, and a linearly interpolated amplitude information, a linearly interpolated position information, and an interpolation unit that provides the linearly interpolated linear prediction coefficient to the synthesis filter, The synthesis filter is based on the linearly interpolated amplitude information, the linearly interpolated position information, and the linearly interpolated linear prediction coefficient. Thus, a speech synthesizer characterized in that a speech waveform having a pitch frequency different from each of the pitch frequencies of the two types of speech waveforms is output as synthesized speech.

Example Next, an example of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram of one embodiment of the present invention. 1 and 2 are audio files storing an audio waveform obtained by analyzing an original audio signal by a pitch interpolation type multipulse method using a representative pitch section pulse train, 1 is an audio file having a low pitch frequency, and 2 is an audio file having a high pitch frequency. This is an audio file. The details of the pitch interpolation type multi-pulse method using the representative pitch section pulse train are omitted since they are described in Reference 3 (Ref. 3; Kazunori Ozawa: "There is little deterioration in sound quality even at 16 kbit / sec or less." Real-time codec using multi-pulse speech coding "Nikkei Electronics 1986.6.1
6 (No. 397) P185-215).

FIG. 2 shows information stored in the audio file.
The pitch period of each frame is 21, PARCOR
A coefficient 22 and sound source information 23 are stored.

FIG. 3 shows the format of the sound source information in the frame. In the case of a voiced section, it is the amplitude information and position information of the pulse train in the representative pitch section, and in the case of the unvoiced section, it is the amplitude information and position information of the pulse train of the entire frame. 31 is the maximum value P _max of the pulse amplitude in the frame, 32 is the amplitude information g _n of the pulse train in the frame, and 33 is the position information _mn of the pulse train in the frame. Here, n corresponds to the number of pulses in the frame.

Returning to FIG. 1, if the pitch period 21 (FIG. 2) of the voice file 1 or 2 is set to 0, the voiced / unvoiced determination circuit 3 determines that the voice file is unvoiced, and The sound source information 9 or 12 is input to the pulse restoration circuit 4.
In the unvoiced section, the unvoiced sound pulse restoration circuit 4 directly inputs the sound source information 9 or 12 to the synthesis fighter 7 because the sound source information 9 or 12 is a pulse train of the entire frame. Voiced / unvoiced judgment circuit 3
If the pitch period 21 (FIG. 2) of the audio file 1 or 2 is set to a value other than 0, the sound file is determined to be voiced, and the sound source information 9 or
Enter 12. In the case of a voiced section, since the sound source information 9 or 12 is a pulse train of the representative pitch section, the voiced pitch control circuit 5 and the pitch periods 10 and
The pitch period interpolation circuit 6 receiving the signal 13 performs the following processing.

First, a method of obtaining the amplitude information g _n and the position information m _n of the pulse train in each frame stored in the audio file 1, 2 4
This will be described with reference to the drawings. FIG. 4 shows a pulse search and interpolation process in a voiced section. In the figure, 41 is the original speech waveform 2
This is the number of frames, and indicates the kth and (k + 1) th frames. As the analysis conditions, the sampling frequency is 8 kHz, and the voice analysis frame length is 20 ms.

42 is a pulse train in the frame. This method divides a frame into sub-frame sections for each pitch period using the pitch period obtained for each frame, interpolates the coefficients of the synthesis filter 7 obtained for each frame for each sub-frame, and filters the sub-frame units. And a predetermined number of pulse trains are calculated for each sub-frame. pulse train of each sub-frame of the k-th frame is _{^{{a 1 K, a 2 K}} , a 3 K, a 4 K, a 5 K}, (k + 1) th pulse sequence of each sub-frame of the frame {a ₁ ^{K + 1} , a ₂ ^{K + 1} , a ₃
^{K + 1} , a ₄ ^{K + 1} , a ₅ ^{K + 1} ｝.

43 is a pulse train in the representative pitch section. In this case, a section in which a good sound similar to the original sound can be reproduced in the entire frame is searched for and selected in several sections. The representative pitch section pulse train of the k-th frame is represented by {a ₂ ^K }, and the representative pitch section pulse train of the (k + 1) -th frame is represented by {a ₂ ^{K + 1} }.

Next, operations of the voiced sound pitch control circuit 5 and the pitch period interpolation circuit 6 will be described with reference to FIGS.

FIG. 54 shows a pulse train in a representative pitch section of a certain frame in an audio file. Here, the number of pulses in one frame will be described as eight. 51 is an audio file having a low pitch frequency (for example, pitch frequency f _a = 150H)
FIG. 9 is a pulse train of a representative pitch section of a certain frame in z). Reference numeral 52 denotes a pulse train of a representative pitch section of a certain frame in an audio file having a high pitch frequency (for example, pitch frequency f _b = 300 Hz), and this frame is a frame corresponding to 51. The pulse amplitude of the 51 pulse train is {g _a1 , g _a2 , ..., g _a8 }, and the pulse position is {m _a1 , m _a2 , ..., m _a8 }
And The pulse amplitude of the 52 pulse train is の g _b1 ,
g _b2 , ..., g _b8 }, and the pulse position is {m _b1 , m _b2 , ..., m _b8 }.

A method for changing the pitch frequency to 200 Hz will be described below.

The method of changing the pitch frequency can be obtained by interpolating the position information and the amplitude information of the pulse trains of the two types of audio files of the pitch frequencies 150 Hz and 300 Hz prepared in advance by the following equations. The pulse amplitude at pitch frequency 200Hz { _gc1 , _gc2 , ..., _gc8 }, and the pulse position is { _mc1 , _mc2 , ..., _mc8 }
Is obtained by the following equation.

Here, n = 1, 2,..., 8.

 FIG. 6 shows a method of interpolating the position information of the pulse train.

61 is an audio file with a low pitch frequency (pitch frequency
a pulse position information of the pulse train of the representative pitch interval of a certain frame in the f _a = 150Hz).

62 is an audio file with a high pitch frequency (pitch frequency
This is pulse position information of a pulse train in a representative pitch section of one frame corresponding to 61 at f _b = 300 Hz).

63 is pulse position information of a pulse train of a representative pitch section of one frame corresponding to 61 when the pitch frequency is set to 200 Hz.

 FIG. 7 shows a method of interpolating the amplitude information of the pulse train.

71 is an audio file with a low pitch frequency (pitch frequency
is a pulse amplitude information of the pulse train of the representative pitch interval of a certain frame in the f _a = 150Hz).

72 is an audio file with a high pitch frequency (pitch frequency
This is pulse amplitude information of a pulse train of a representative pitch section of one frame corresponding to 71 at (f _b = 300 Hz).

Reference numeral 73 denotes pulse amplitude position information of a pulse train of a representative pitch section of one frame corresponding to 71 when the pitch frequency is set to 200 Hz.

As described above, the pitch frequency can be changed as described.

In this way, the pitch period interpolation circuit 6, the audio file 1 that is known beforehand that when the f _c = 200 Hz pitch control information 17 are input f _a = 150 Hz and the audio file 2 f
_b = 300 Hz is supplied to the pitch control circuit 5, and the pulse amplitude g _cn and the pulse position m _cn are obtained by the equations (1) and (2).

The linear prediction coefficient interpolation circuit 8 receives the PARCOR coefficients 11 and 14 of the audio file 1 and the audio file 2, performs linear interpolation, and inputs the linear interpolation to the synthesis filter 7. The synthesis filter 7 is a sound source restoration unit
The synthesized speech 15 is output based on the sound source information input from 16 and the PARCOR coefficient input from the linear prediction coefficient interpolation circuit 8.

[Effects of the Invention] As described above, the present invention relates to a speech synthesizer using a speech waveform coding method (multi-pulse method) having a speech generation model, which has two different pitch frequencies analyzed by the multi-pulse method. The pitch is continuously changed by a simple pitch control method in which the pitch period is changed by means of linearly interpolating the amplitude information and position information of the pulse train of the representative pitch section in the frame of the type of speech waveform and linearly interpolating the linear prediction coefficient. Can be varied.

Therefore, the reference sound is converted to a sound element piece by the multi-pulse method.
If the waveform data is prepared by decomposing it into synthesis units such as CV-VC phoneme chains, the pitch can be smoothly changed by using the pitch control of this method during rule synthesis, and a rule-synthesized speech with naturalness can be obtained. There is an effect that can be.

In addition, since the speech waveform encoding method uses the waveform itself, several steps (5 steps) of waveform data with different pitch periods are used.
In the present invention, two types of voice waveforms having different pitch frequencies (a voice waveform having a high pitch frequency and a voice waveform having a low pitch frequency) are required.
And the memory capacity can be reduced. In addition, since speech can be compressed by the multi-pulse method, a compact speech synthesizer can be realized.

[Brief description of the drawings]

FIG. 1 is a block diagram of a speech synthesizer according to an embodiment of the present invention. FIG. 2 is a diagram showing a format of information stored in a speech file of the speech synthesizer of FIG.
The figure shows the format of sound source information in a frame. FIG. 4 is a processing waveform showing a processing method, and FIG. 5 is a pulse train of a representative pitch section of a certain frame in an audio file. FIG. 6 shows a method of interpolating the position information of the pulse train, and FIG. 7 shows a method of interpolating the amplitude information of the pulse train. 1 ... voice file (voice file with low pitch frequency), 2 ... voice file (voice file with high pitch frequency), 3 ... voiced / unvoiced judgment circuit, 4 ... unvoiced sound pulse restoration circuit, 5 ... voiced sound Pitch control circuit, 6 ... Pitch period interpolation circuit, 7 ... Synthesis filter, 8 ... Linear prediction coefficient interpolation circuit, 9 ... Sound source information (audio file with low pitch frequency), 10 ... Pitch period (Pitch frequency PARCOR coefficient (sound file with low pitch frequency), 12 ... Sound source information (sound file with high pitch frequency), 13 ... Pitch cycle (sound file with high pitch frequency), 14 ... PARCOR coefficient (sound file with high pitch frequency), 15 synthesized speech, 16 sound source pulse restoration unit, 17 pitch control information, 21 pitch cycle in frame, 23 sound source information in frame, 31 ... Pal in the frame The maximum value of the pulse amplitude, 32 ... the amplitude of the pulse train in the frame,
33: Position of pulse train in frame, 41: Two frames of original voice waveform, 42: Pulse train in frame, 43:
… Pulse train in representative pitch section, 51 …… Pulse train in representative pitch section (audio file with low pitch frequency), 52…
Pulse train of representative pitch section (voice file with high pitch frequency), 61 ... Pulse train of representative pitch section (voice file with low pitch frequency), 62 ... Pulse train of representative pitch section (voice file with high pitch frequency), 63 ... … Pulse train of representative pitch section (interpolated pitch frequency), 71
... Pulse train of the representative pitch section (sound file with low pitch frequency), 72... Pulse train of the representative pitch section (sound file with high pitch frequency), 73... Pulse train of the representative pitch section (interpolated pitch frequency).

Claims (1)

(57) [Claims] 1. A speech synthesizer using a multi-pulse method, in which an input speech is divided into frames of a fixed time length and a pitch is extracted from the input speech and represented by a fixed number of driving sound source pulses for each pitch. Linearly interpolating amplitude information, position information, and a linear prediction coefficient of a pulse train in a representative pitch section in frames of two types of speech waveforms having different pitch frequencies analyzed by the multi-pulse method, and linearly interpolated amplitude information; Interpolating means for providing an interpolated position information and a linearly interpolated linear prediction coefficient to a synthesis filter, wherein the synthesis filter includes the linearly interpolated amplitude information, the linearly interpolated position information, and the linear interpolation A voice waveform having a pitch frequency different from each of the pitch frequencies of the two types of voice waveforms based on the obtained linear prediction coefficients; Speech synthesis apparatus characterized by and output.