JP3270869B2 - Pitch converter - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pitch converter for converting the pitch of a voice, and more particularly to a pitch converter which can easily convert the pitch of a voice without impairing the naturalness of the voice and the personal characteristics of the speaker. The present invention relates to a pitch conversion device.

[0002]

2. Description of the Related Art Various apparatuses and methods for converting the pitch (basic period, pitch) of speech have already been proposed. As a conventional simple pitch conversion device, a voice input from a microphone is converted into a digital signal by an A / D converter, and this digital voice signal is converted into a predetermined period (time interval).
And reads the digital audio signal written in the memory at a cycle different from the cycle of writing,
2. Description of the Related Art There is known a pitch conversion device that returns an analog audio signal by a D / A converter, amplifies the signal by an amplifier, and outputs the signal from a speaker. In other words, in this pitch conversion device, the pitch is converted by expanding and contracting the audio data on the time axis in accordance with the ratio of the period of writing the audio signal to the memory and the period of reading the audio signal from the memory. However, when pitch conversion is performed using this pitch conversion device, the frequency amplitude characteristics are distorted, and the pitch-converted voice output from the speaker loses the naturalness and personal characteristics of the original voice. I'm having problems with unnatural audio. When the degree of pitch conversion is increased, the unnaturalness becomes more pronounced.

Various pitch conversion methods have been proposed to improve such unnaturalness. As such a conventional example, for example, NHK Giken R & D, No. 19, 1992,
May, as disclosed on pages 11-24, FIG.
The “linear prediction method” illustrated in (b) and (c) is known. This method separates the input voice into vocal tract characteristics and sound source characteristics, and cuts off the residual waveform to raise the voice, or performs zeroing to lower the voice, as illustrated in FIG. 5 (c). , And drives the vocal tract filter shown in FIG. 5B, and is also referred to as “residual drive type analysis / synthesis method”. However, in this linear prediction method, since the separation between the sound source information and the vocal tract information is incomplete, the residual waveform contains a considerable amount of components due to the vocal tract information other than the sound source information, and the pitch period is reduced. When changing, these spectra are also deformed, which has the problem that the quality is directly reduced. Such degradation in quality due to imperfect separation is more pronounced in the analysis and synthesis of relatively high-pitch female or child voices than male voices.

A method and apparatus for improving the above-described linear prediction method are also known (see NHK Giken R & D).
D, No. 19, 1992, May, 11-24. FIG. 3 shows the configuration of the apparatus. The pitch conversion apparatus shown in FIG. 3 is a pre-processing unit that determines whether an input voice is silent, unvoiced, or voiced, and determines a pitch. The voice censor converts a voice waveform into a pulse train (sound source) every pitch cycle using a complex cepstrum. Information) and a unit response (vocal tract information), a conversion and synthesis unit that adds a desired pitch change to the pitch pulse train and convolves the pitch pulse and the unit response for each pitch cycle based on the result to synthesize a speech waveform And a post-processing unit that removes non-harmonic components generated by the pitch change and adjusts the gain, and connects the obtained waveform to the silent and unvoiced sections recorded in the pre-processing unit to generate a final audio output. And The pitch conversion apparatus shown in FIG. 3 is intended to separate the vocal tract information from the vocal tract information in such a manner that the sound source information is as pulse-like as possible in order to improve the above-described problem of the linear prediction method. For this reason, the speech waveform is separated into a pulse train and a unit response by complex cepstrum analysis, and the pulse train is zero-phased to sharpen the pulse train. Thus, sound source information (pulse train) and vocal tract information (unit response) are clearly separated without loss of information.

FIG. 4 is a diagram showing a more detailed processing form of the analysis unit described above. The analysis unit includes a complex cepstrum analysis process (step "1"), a separation process using a comb lifter (step "2"), a time waveform reproduction process (step "3"), zero-phase pulse train, and inverse compensation of unit response. Processing (step "4") and recording processing of the pitch pulse and the unit response (step "5") are performed.

[0006] In the complex cepstrum analysis process in step 1, the target pitch section in the voiced section is 2.5 times the pitch section length from the position temporally preceding by one pitch section length. , A waveform is cut out by applying a Hanning window, and the number of analysis points is increased by performing zeroing in the latter half of the cut out waveform, and a complex cepstrum is obtained by performing an inverse fast Fourier transform (inverse FFT or IFFT). In other words, as a method of obtaining this complex cepstrum, a zero sequence is added to the speech waveform cut out by the Hanning window to increase the number of analysis points, inverse FFT is performed, and the phase difference between two adjacent sample values is reduced. When this phase difference is larger than π (radian), 2
A “phase unwrapping method” is used in which phases are connected by adding or subtracting π (radian). When a cepstrum component is separated by a lifter to be described later, a linear phase component included in the unwrapped phase adversely affects the reproduced waveform and may be perceived as distortion. After removing and separating with lifter,
When reproducing the time waveform, a process of compensating again for the portion removed in the pulse train is performed.

[0007] The lifter means a filter in the keflesin region. Separation by the comb lifter in step 2
This is a process in which a cepstrum is separated into a component that is an integral multiple of the pitch period and other components using a comb-shaped lifter and an inverted comb-shaped lifter that performs the reverse process. The reproduction of the time waveform in step 3 is processing for obtaining the time waveform of each of the separated cepstrum. By convolving the separated pulse train with the unit response, the original cut out waveform is obtained.

The step 4 of zeroing the pulse train and the inverse compensation of the unit response will be described. Since the position and amplitude at which each pulse of the pulse train returned to the time waveform appears in the frame is significantly different even if the position or length of the analysis window (Hanning window) for the speech waveform is slightly changed, analysis and synthesis are performed by pitch synchronization. Is difficult to do. Therefore, pitch synchronization is performed while determining the position of the pulse with respect to the original speech waveform, and this pulse is further sharpened to reduce the effect of truncation at the time of pitch change and the effect of spectral deformation due to zero padding. Perform the conversion process. For the unit response, inverse compensation of the zero-phase processing is performed on the unit response obtained for each frame so that the original cut-out waveform is obtained when the zero-phased pulse train is convolved.

In step 5, the pitch pulse and the unit response are recorded by applying a Hanning window of a pitch section length centering on the pulse to cut out only the maximum pulse from the frame of the zero-phased pulse train. This is a process of recording together with the inversely compensated unit response as the driving sound source representative of.

The analysis section is moved, and the above-described step 1 is performed.
Step 5 is repeated. This operation results in a zero-phased pitch pulse for each frame and a corresponding decompensated unit response. By arranging the zero-phased pitch pulses for each frame, a pitch pulse train corresponding to the original speech can be formed.

FIG. 4 shows a more detailed processing form of the conversion / synthesis unit shown in FIG. In Step 1, each pitch section of the obtained pulse train is changed to a desired cycle.
To lower the original voice to a lower voice (lower the pitch), the pitch interval is extended by performing zero rounding between the pitch pulses. Conversely, when raising the pitch, the data between the pitch pulses is censored and the pitch interval is reduced. To shorten. In step 2, pitch pulses are thinned out or repeated for each pitch section from the pitch pulse train, and a desired change is made to the time length. In step 3, the pitch pulse representing each frame and the corresponding unit response are convolved for each pitch section. Thereby, the obtained waveform becomes a time waveform for one pitch representing each pitch section. In step 4, superposition is performed based on the positions of the respective pulses in the obtained pitch pulse train.

FIG. 5 (a) is the same as FIG.
FIG. 3 is a schematic diagram of the analysis unit shown in FIG. FIG. 5B is a frequency characteristic diagram of the vocal tract filter of the above-described linear prediction method, FIG. 5C is a graph showing a pitch change of the linear prediction method, and FIG.
A graph showing pitch change by pulse train processing shown in
(E) is a frequency characteristic diagram of a unit response by the processing shown in FIG. 3. The method applying the complex cepstrum analysis can clearly separate the sound source information (pulse train) and the vocal tract information (unit response) without loss of information as compared with the linear prediction method described above, and can reduce the characteristics of the original speech. The effect is that the pitch can be converted while maintaining the pitch.

[0013]

However, FIGS.
The conventional pitch conversion device described with reference to FIG. 5 needs to perform phase adjustment when convolving the sound source waveform and the time response of the vocal tract component, and since this phase adjustment is complicated, the entire pitch conversion device has A problem of complicated circuit configuration has been encountered. In addition, it takes a long time to adjust the phase, so that the processing time is long and the responsiveness is low.

The present invention solves the above-mentioned problems in the conventional pitch conversion device, and enables pitch conversion in a short time with a relatively simple circuit configuration, and also enables pitch conversion while maintaining more naturalness. It is an object of the present invention to provide a pitch conversion device.

[0015]

SUMMARY OF THE INVENTION In the present invention, in order to solve the above-mentioned problems and achieve the above-mentioned object, the first conventional simple pitch converting device described above and the last conventional pitch converting device described above are used. Taking advantage of the above, a pitch conversion device that compromises between them is constructed. Therefore, according to the present invention, means for changing the pitch of the sound while preserving the phase of the original sound signal, means for extracting a sound source characteristic signal from the sound signal with the changed pitch, A pitch conversion device is provided, comprising: means for extracting a vocal tract characteristic signal from a voice signal; and means for performing voice synthesis using the extracted vocal tract characteristic signal and the extracted sound source characteristic signal.

[0016]

First, the pitch of the voice is changed while the original voice is preserved, as in the method first described as the prior art. As in the last-mentioned method of the prior art, the sound source characteristics are extracted from the pitch-changed audio signal and the vocal tract characteristics are extracted from the original audio signal. These extracted sound source characteristic signal and vocal tract characteristic signal are subjected to speech synthesis.
Since the pitch of the extracted sound source characteristic signal has already been changed, there is no need to adjust the phase with the vocal tract characteristic signal. on the other hand,
Since the sound source characteristic signal and the vocal tract characteristic signal are separated, the sound source characteristic signal and the vocal tract characteristic signal are clearly separated, so that a sound with a changed pitch can be obtained while maintaining the characteristics of the original sound. .

[0017] Preferably, the changing of the pitch of the voice is performed by changing a cycle of writing the voice signal and a cycle of reading the voice signal,
Change the pitch. Preferably, a complex cepstrum analysis is performed for the extraction of the sound source characteristic signal and the extraction of the vocal tract characteristic signal. More preferably, the extracted sound source characteristic signal is multiplied by the extracted claim characteristic signal, the result of the multiplication is inverse Fourier-transformed, and the Fourier-transformed signal is overlap-added to perform speech synthesis.

[0018]

FIG. 1 is a block diagram showing an embodiment of a pitch conversion apparatus according to the present invention.
Shown in The pitch converter shown in FIG. 1 includes a microphone 1 for recording voice, an A / D converter 2 for converting the recorded microphone-recorded voice signal S1 into a digital signal,
An original voice signal pitch changing device 20 for performing pitch conversion on the original digital voice signal S2, and a first vocal tract characteristic / for extracting a vocal tract characteristic signal from the original digital voice signal S2 /
A sound source characteristic extraction device 6, a second vocal tract characteristic / sound source characteristic extraction device 7 for extracting a sound source characteristic signal from the pitch-converted digital audio signal S5, and an extracted vocal tract characteristic signal S6B and an extracted sound source characteristic signal S7A. A vocal tract / sound source synthesizer 30 for synthesis, a D / A converter 11 for converting a synthesized voice signal into an analog signal, an amplifier 12, and a speaker 13
Having.

The A / D converter 2 is, for example, 12 KH_Z
To convert the microphone-recorded audio signal S1 into the original digital sound
It is converted to a voice signal S2. Original voice signal pitch changing device 2
0 is the memory writing device 3, the memory 4, and the memo
It comprises a rereading device 5. Memory 4 is a ring
A semiconductor memory having a buffer configuration, for example, a DRAM.
You. The memory writing device 3 stores the first cycle
12KH as_ZOutput from the A / D converter 2 at a constant cycle of
The input original digital audio signal S2 is written. Me
The memory reading device 5 is, for example, 6 to 24 KH_ZRange of
Is written to the memory 4 at a cycle (frequency) that varies with
Read out the original digital audio signal and read the digital audio signal
Read as S4, cycle specified by read frequency
To output as digital voice signal S5 after pitch conversion.
You. This reading cycle is determined according to the pitch to be converted.
It is. 6-24KH above_ZThe cycle that changes in the range of
Injection cycle 12KH_ZOctave of 1/2 to 2 times
Indicates the range in which the pitch (pitch) can be converted by the pitch. for example
For example, if you change the pitch to 1/2 octave, read
The cycle is 6KH _ZIs set to In this case, memory 4
The original digital audio signal that has been decimated one by one
Be counted. On the other hand, when converting the pitch to double octave
Means that the read cycle is 24KH_ZIs set to This place
The original digital voice signal written from the memory 4
Signal is read, and the two original data
Digital audio signal interpolation, for example, linear interpolation
To compensate for the missing audio signal. In this pitch conversion
Note that the phase of the original pitch signal is preserved.
I want to be.

The first vocal tract characteristic / sound source characteristic extracting device 6
Finally, a complex cepstrum analysis is performed as in the prior art pitch converter described above, and the original digital speech signal S
2, an extracted sound source characteristic signal (pulse train information) S6A and an extracted vocal tract characteristic signal (formant: unit response information) S6B are separated and extracted. However, the extracted sound source characteristic signal S6A
Are not used, and only the extracted vocal tract characteristic signal S6B is used. It should be noted that the phase of the extracted vocal tract characteristic signal S6B is also preserved. Therefore, the first vocal tract characteristic / sound source characteristic extraction device 6 is not provided with the phase adjustment circuit in the complex cepstrum analysis described as the related art.
In addition, as a method of the complex cepstrum analysis, as described above, in addition to the “phase unwrapping method”, a method that does not use the phase unwrapping, or a signal having a general convolution relation is subjected to a special conversion. Homomorphic signal processing that decomposes by filtering or the like can be applied.

The second vocal tract characteristic / sound source characteristic extracting device 7
It has the same circuit configuration as the above-described first vocal tract characteristic / sound source characteristic extraction device 6, and separates the extracted vocal tract characteristic signal S7A and the extracted vocal tract characteristic signal S7B from the digital voice signal S5 after pitch conversion. To extract. Extracted vocal tract characteristic signal S7
B is not used, and only the extracted sound source characteristic signal S7A is used. As described above, the phase of the original digital audio signal S5 after pitch conversion is preserved, so that the phase of the extracted sound source characteristic signal S7A is also preserved.

The vocal tract / sound source synthesizer 30 comprises a multiplier 8, an inverse FFT unit 9, and an overlap adder 10. The multiplier 8 includes an extracted vocal tract characteristic signal S6B and an extracted sound source characteristic signal S6B.
7A and outputs a sound source / vocal tract multiplication signal S8.
The inverse FFT device 9 performs an inverse Fourier transform on the sound source / vocal tract multiplication signal S8 at high speed, and generates a sound source characteristic signal (pulse train).
To compensate. Although the inverse FFT (IFFT) is used to perform the inverse Fourier transform at high speed, the inverse Fourier transform may be performed in principle. The overlap adder 10 overlap-adds the inverse Fourier transform signal S9, that is, overlaps the inverse Fourier transform signal S9 on a frame-by-frame basis to reproduce a pitch-converted audio signal.

The D / A converter 11 converts the overlap addition signal S10 into an analog signal. Pitch-converted analog audio signal S
11 is amplified by an amplifier 12 with a predetermined gain, and an amplified pitch-converted analog audio signal S12 is output from a speaker 13.

As described above, in the above-described embodiment, after the pitch of the original digital audio signal S2 is converted by the original audio signal pitch changing device 20 while the phase is preserved, the second vocal tract is changed. The characteristic / sound source characteristic extraction device 7 extracts the extracted sound source characteristic signal S7A, and the first digital vocal tract characteristic / sound source characteristic extraction device 6 extracts the extracted vocal tract characteristics in a state where the phase of the original digital audio signal S2 is preserved. Since the signal S6B is extracted, it is not necessary to adjust the phase when synthesizing the extracted vocal tract characteristic signal S6B and the extracted sound source characteristic signal S7A. Therefore, pitch conversion can be realized in a short time with a simple circuit configuration. Moreover, since the sound source characteristics and the vocal tract characteristics are clearly separated and extracted by the complex cepstrum analysis, the pitch can be converted while maintaining the naturalness of the original voice and the characteristics of the individual.

The embodiment of the present invention is not limited to the above-described embodiment, but may take other various forms. For example, the above-described write cycle (frequency) and read cycle are merely examples. As the method of interpolation in the memory reading device 5, in addition to simple linear interpolation,
Various interpolations, such as weighted interpolation, can be applied. In implementing the present invention, known circuits can be used for the first vocal tract characteristic / sound source characteristic extracting device 6, the second vocal tract characteristic / sound source characteristic extracting device 7, the inverse FFT device 9, the overlap adder 10, and the like. In particular, the detailed circuit configuration was not specified. Therefore, various complex cepstrum analysis circuits, inverse FFT devices, overlap adders, and the like applied to signal processing can be applied.

[0026]

According to the present invention, the original voice sound naturalness and personal characteristics can be maintained in a short time with a simple circuit configuration.
The pitch of an audio signal can be arbitrarily changed.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of an embodiment of a pitch conversion device of the present invention.

FIG. 2 is a conceptual configuration diagram of a pitch conversion device.

FIG. 3 is a first partial configuration diagram of a conventional pitch conversion device.

FIG. 4 is a second partial configuration diagram of a conventional pitch conversion device.

5A and 5B are diagrams showing a partial configuration of a conventional pitch conversion device and processing characteristics thereof, and FIG. 5A is a diagram showing FIG. 2 described with reference to FIG.
, (B) a frequency characteristic diagram of a vocal tract filter of the linear prediction method, (c) is a graph showing a pitch change of the linear prediction method, and (d) is based on the pulse train processing shown in FIG. FIG. 4E is a graph showing a pitch change, and FIG. 4E is a frequency characteristic diagram of a unit response by the processing shown in FIG.

[Explanation of symbols]

1. Microphone 2. A / D converter 3. Memory writing device 4. Memory 5. Memory reading device 6. First vocal tract characteristic / sound source characteristic extraction device 7. Second voice Road characteristic / sound source characteristic extraction device 8. Multiplier 9. Inverse FFT device 10. Duplicate adder 11. D / A converter 12. Amplifier 13. Speaker 20. Original voice signal pitch changing device 30 vocal tract / sound source synthesizer S1 microphone sound signal S2 original digital audio signal S3 original digital audio signal for writing S4 read digital audio signal S5 digital after pitch conversion Voice signal S6A ... Extracted sound source characteristic signal S6B ... Extracted vocal tract characteristic signal S7A ... Extracted sound source characteristic signal S7B ... Extracted vocal tract characteristic signal S8 ... Sound source / vocal tract multiplication signal S9 ... Reverse Fourier transform signal S10: overlapping addition signal S11: pitch-converted analog voice signal S12: amplified pitch-converted analog voice signal

Continuation of the front page (56) References JP-A-3-259196 (JP, A) JP-A-3-18900 (JP, A) JP-A-2-137900 (JP, A) JP-A-58-43498 (JP) JP-A-51-83504 (JP, A) JP-A-3-51899 (JP, A) JP-A-1-147600 (JP, A) JP-A-2-134699 (JP, A) 1-304500 (JP, A) JP-A-1-302299 (JP, A) JP-A-4-37999 (JP, U) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 21/04

Claims (5)

(57) [Claims] A means for changing a pitch of the sound while preserving the phase of the original sound signal; a means for extracting a sound source characteristic signal from the sound signal having the changed pitch; A pitch conversion device comprising: means for extracting a vocal tract characteristic signal; and means for performing speech synthesis using the extracted vocal tract characteristic signal and the extracted sound source characteristic signal. 2. The memory according to claim 1, wherein said means for changing the pitch of the voice includes a memory, a writing means for writing the original voice signal in the memory at a first cycle, and a voice signal written to the memory for a second cycle. 2. The pitch conversion device according to claim 1, further comprising: a reading unit that reads the audio data by using a ratio between a first period and a second period. 3. The pitch conversion device according to claim 1, wherein the means for extracting the sound source characteristic signal performs a complex cepstrum analysis to extract the sound source characteristic signal. 4. The pitch conversion device according to claim 3, wherein said means for extracting a claim signal performs a complex cepstrum analysis to extract the claim signal. 5. A speech synthesizer, comprising: a multiplier for multiplying the extracted sound source characteristic signal by the extracted claim characteristic signal; a circuit for performing an inverse Fourier transform of a multiplication result of the multiplier; 5. The pitch conversion device according to claim 4, further comprising a circuit for overlappingly adding the converted signals.