JP3125937B2 - Voice pitch conversion method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice pitch conversion method for an apparatus for performing pitch control, and more particularly, to a case where an erroneous peak position is given in consideration of both a peak position of an input voice and a pitch period to be converted. The present invention relates to a voice pitch conversion method capable of coping with the above.

[0002]

2. Description of the Related Art Conventionally, in a voice pitch converter for performing voice pitch control, as shown in FIG. 2, when a voice to be converted is input from an input terminal 21 to a processing device, a pre-processing unit 22 receives an analog signal. / Perform pre-processing such as digital conversion.
On the other hand, the local peak position 23 in the audio waveform which is temporally synchronized with the data is obtained by visual observation (or a method of automatic peak position extraction). These data and the peak position are input to the waveform cutout / superposition unit 24 and stored. Further, a pitch pattern 25 that is a target of voice pitch conversion is input to the waveform cutout / superposition unit 24. Thereby, based on the accumulated local peak position, the waveform is multiplied by a time window function such as a hanning window that matches the pitch period of the synthesis target with the audio waveform centering on the local peak position. The waveform is cut out and the waveform is superimposed again according to the target pitch pattern 25. In this way, a synthesized speech is formed by the waveform cutout / superposition unit 24, post-processing such as digital / analog conversion is performed on the data by the post-processing unit 26, and a converted speech is obtained from the output terminal 27. In addition, about the conventional pitch control method, for example,
"Hirokawa, Hakoda, Pitch Control Method in Waveform Editing Rule Synthesis Method, Proceedings of the Acoustical Society of Japan, March 1990-
4-7 ".

[0003]

In the above-mentioned prior art, when applying a window function for extracting a waveform, the waveform is extracted using a window function that only takes into account that the window function matches a target pitch. Therefore, no consideration is given to the pitch structure of the input speech waveform, and the pitch structure of the input speech waveform may be damaged. In the process of automatically extracting the local peak position, for example, if there is a peak extraction error due to the fact that the input voice does not have a pitch structure, if the waveform is cut out using a window function according to a conventional method, It is expected to cause deterioration such as noise on voice. An object of the present invention is to improve such a problem by determining the shape of a window function for waveform extraction in consideration of both an input speech waveform and a target pitch pattern, thereby improving a given peak position. It is an object of the present invention to provide a voice pitch conversion method capable of absorbing high-quality errors and forming a high-quality synthesized voice.

[0004]

In order to achieve the above object, a voice pitch conversion method according to the present invention is characterized in that, in a waveform cutout / superposition unit, a time interval of a local peak position of an input voice extracted by a peak position extraction unit is extracted. The distance (wlf, wlb) between the peak position to which the window function is applied and the preceding and succeeding peak positions in time series with the peak position is obtained, and the coefficient C obtained from the conversion rate R to the target pitch is calculated for the distance. , The window length is determined by multiplying by (1) to generate a left-right asymmetric window function (Equation (4)) considering both the input voice and the converted voice.

[0005]

In the present invention, when performing waveform superposition in the waveform cutout / superposition unit, the distance between the local peak position to which the time window function is applied and the preceding and succeeding peak positions in time series with the peak position is determined. The window length is determined by multiplying those distances by a coefficient obtained from the conversion rate to the target pitch. That is, the window length of the left (in time series) and right (in time series) portions of the inflection point (peak position) in the section where the window function is applied is twice as long as the immediately preceding peak position. hann
Since they are equivalent to the left and right halves of the window function such as the ing window, respectively, by combining both sides to form a left-right asymmetric window function, both input voice and converted voice are taken into account, The shape of the window function can be determined. This makes it possible to absorb an error at a given peak position while maintaining the pitch pattern of the original voice, and thus it is possible to obtain a high-quality synthesized voice at a waveform level.

[0006]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 3 is a configuration diagram showing a part of the voice pitch conversion device in one embodiment of the present invention, and FIG. 4 is an example diagram of peak positions in a voice waveform having a pitch structure in one embodiment of the present invention. The voice pitch conversion device of the present embodiment includes an input device for inputting voice, an output device for outputting a voice pitch conversion result of the input voice, data after processing, a target pitch pattern for voice pitch conversion, and the like. It comprises an external storage device for storing, a processing device shown in FIG. 3, and the like. This processing device includes a CPU, a memory, and the like, and includes a pre-processing unit 32, a peak position extracting unit 33, a waveform cutout / overlapping unit 34, and a post-processing unit 36, as shown in FIG. With such a configuration, when performing voice pitch conversion, a voice signal is input from the input terminal 31 and is input to the preprocessing unit 32. The preprocessing unit 32 performs analog / digital conversion through a low-pass filter, converts the voice into digital data, and sends the digital data to the peak position extraction unit 33. The peak position extraction unit 33 extracts a peak position from the audio waveform, and inputs information on the peak position to the waveform cutout / superposition unit 34. This peak position is shown in FIG.
Are the portions indicated by the “*” mark. The waveform cutout / superposition unit 34 stores these data, inputs the target pitch pattern 35, cuts out the audio waveform based on the peak position from the peak position extraction unit 33, and becomes the target pitch pattern. The digital data of the synthesized voice obtained is input to the post-processing unit 36 by overlapping the waveforms as described above. The post-processing unit 36 performs digital-to-analog conversion of the digital data, and outputs a converted voice to an output terminal 37 through a low-pass filter.

Here, the waveform cutting / overlapping section 34
Is described in detail. FIG. 5 is a flowchart showing the processing of the waveform cutout / superposition unit in one embodiment of the present invention. The conversion rate of the conversion target or the target pitch pattern 35 is input to the waveform cutout / superposition unit 34 of the present embodiment together with the information on the peak position (501, 502). The waveform superimposition position is determined from the input peak position sequence and the conversion target (50).
3) A cutout position corresponding to the overlapping position is determined (504). Next, a window function for waveform extraction is generated in consideration of the environment of the waveform extraction position (505).
Then, a speech waveform is cut out using the generated window function (506), and a superposition process is performed on the previously determined superposition position (507). These processes are sequentially repeated until the synthesis into the target pitch pattern is completed (508), thereby forming digital data of the synthesized voice.
Input to 6.

Next, the window function generation processing of step 505 in FIG. 5 will be described in detail. FIG. 1 is a flowchart showing a process of generating a window function for extracting a waveform in one embodiment of the present invention. FIG. 6 is an explanatory diagram of a left-right asymmetric window in one embodiment of the present invention. FIG. 8 is an explanatory diagram when the pitch period is shortened by the window function, and FIG. 8 is an explanatory diagram when the pitch period is increased by the window function according to the embodiment of the present invention. In FIG. 6, 61 is the peak position immediately before the peak position 62 of interest, 62 is the peak position of interest, 63 is the peak position after the peak position 62 of interest, 64 is the window function, and 65 is one. Previous peak position 6
A distance (wlf) to 1 and a distance 66 to the next peak position 63 (wlb). In the present embodiment, FIGS.
As described above, the window length (wlf + wlb) which is the outline of the window function is determined from the peak positions 61 and 63 before and after the time series of the given waveform cutout position (101). If the peak positions before and after in the time series have the general length of the pitch period, the processing of the present embodiment is continued as it is. Perform processing to assign fixed values. Next, for the obtained window length, the conversion rate of the period is calculated from the conversion target (10
2), the window length is set to a value suitable for the conversion target (10)
3). Further, a window function is generated from the obtained window length (104, 105). Here, the definition formula of the window function of the present embodiment will be described. First, the following equation (1) shows the definition equation of the conventional window function fw (n).

(Equation 1) Where wlf: Number of samples up to the previous peak position wlb: Number of samples up to the next peak position. In the present embodiment, when the pitch frequency is increased (short period), the window length is shortened, and the window frequency is decreased (long period). In this case, a process for increasing the window length is performed. That is, using the coefficient C obtained from the frequency conversion rate, the equation (1)
To transform. This C is expressed by the following equation (2).

(Equation 2) Note that the frequency conversion rate is R [%] (R> 0). By applying C, the relationship between the frequency f and the converted frequency f ′ and the relationship between the cycle T and the converted cycle T ′ are as shown in the following equation (3).

(Equation 3) Next, the window function fw (n) applied to the present embodiment is expressed by the following equation (4).
Shown in

(Equation 4) Where wlf: Number of samples up to the previous peak position wlb: Number of samples up to the next peak position. In the definition equation (1), the window function for waveform extraction is configured based on the hanning window. Regarding the window function serving as a base, even when a hamming window, a blackman window, or the like is applied, the configuration of the asymmetric window function of the present embodiment can be achieved by using the phase portion used in the definition equation (4). . An example in which the pitch period is shortened by such an asymmetric window function is shown in FIG. In this case, the input voice 7
The peak position at 1 (the part marked with *) is the synthesized voice 72
Then, it moves to a position matching the target pitch pattern (indicated by a triangle), and it can be seen that the interval between the peak positions is reduced. FIG. 8 shows an example of increasing the pitch period. In this case, the peak position in the input sound 81 (the part marked with *)
Indicates that the synthesized voice 82 moves to a position matching the target pitch pattern (the portion indicated by the triangle), and it can be seen that the interval between the peak positions is extended. In the present embodiment, based on the definition equation (4), a speech waveform is cut out using the window functions generated in steps 104 and 105, and a superimposition process is performed on the previously determined superposition position, thereby obtaining the pitch of the original voice. A synthesized speech with a preserved structure can be obtained.

[0009]

According to the present invention, the shape of the window function for waveform extraction is determined by considering both the input speech waveform and the target pitch pattern, thereby preserving and providing the pitch pattern of the original speech. It is possible to absorb the error of the peak position thus obtained, and it is possible to form a high-quality synthesized speech at the waveform level.

[0010]

[Brief description of the drawings]

FIG. 1 is a flowchart illustrating a process of generating a window function for waveform cutout according to an embodiment of the present invention.

FIG. 2 is a configuration diagram showing a part of a conventional voice pitch conversion device.

FIG. 3 is a configuration diagram showing a part of a voice pitch conversion device according to an embodiment of the present invention.

FIG. 4 is a diagram illustrating an example of a peak position in a voice waveform having a pitch structure according to an embodiment of the present invention.

FIG. 5 is a flowchart illustrating a process of a waveform cutout / superposition unit according to an embodiment of the present invention.

FIG. 6 is an explanatory diagram of a left-right asymmetric window in one embodiment of the present invention.

FIG. 7 is an explanatory diagram of a case where a pitch period is shortened by a window function according to an embodiment of the present invention.

FIG. 8 is an explanatory diagram of a case where a pitch period is lengthened by a window function according to an embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 21 Input terminal 22 Pre-processing part 23 Peak position 24 Waveform extraction / superposition part 25 Target pitch pattern 26 Post-processing part 27 Output terminal 31 Input terminal 32 Preprocessing part 33 Peak position extraction part 34 Waveform extraction / superposition part 35 Target pitch Pattern 36 Post-processing unit 37 Output terminal 61 Previous peak position 62 Peak position of interest 63 Next peak position 64 Window function 65 Distance to previous peak position (wlf) 66 Next peak position Distance to (wlb) 71 Input voice 72 Synthetic voice 81 Input voice 82 Synthetic voice

Continuation of the front page (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 21/00-21/04 G10H 1/043

Claims (1)

(57) [Claims] An audio waveform is cut out by a time window function based on a local peak position in an input audio waveform.
In the method of performing voice pitch conversion to a target pitch pattern by superimposing the waveforms again, the local peak position includes a peak position to which a time window function is applied, and a peak position before and after the time position adjacent to the peak position. A voice pitch conversion method, wherein a window length is obtained by calculating a distance from a peak position of the target pitch and multiplying the distance by a coefficient obtained from a conversion rate to a target pitch.