JP5191459B2 - Vocal tract cross-section function extraction apparatus and method - Google Patents

Description

  The present invention relates to a vocal tract cross-sectional area function extraction device and an extraction method thereof.

  Non-Patent Document 1 discloses a conventional method for extracting a vocal tract cross-sectional area function. In this vocal tract cross-sectional area function extraction method, the frequency characteristics of the voice data are flattened by removing the radiation characteristics of the vocal cord sound source and lips not included in the characteristics of the vocal tract by the adaptive inverse filter method. For this reason, in the conventional method for extracting the vocal tract cross-sectional area function, it is possible to extract the vocal tract cross-sectional area function after removing the vocal cord sound source and the radiation characteristic in advance.

Takayuki Nakajima et al., "Estimation of vocal tract shape by deconvolution and adaptive speech analysis system", Journal of the Acoustical Society of Japan, Vol. 34, No. 3 (1978), P157-P166

  However, the conventional method for extracting the vocal tract cross-sectional area function does not take into account a voice signal having a high sampling frequency required for synthesizing a high quality voice. Experimentally, it has been confirmed that a voice signal having a high sampling frequency is deteriorated during voice quality control.

  The present invention has been made in view of the above-described conventional situation, and provides a vocal tract cross-sectional area function extraction apparatus and an extraction method thereof that can satisfactorily control voice quality even with a high sampling frequency audio signal. Is a problem to be solved.

The vocal tract cross-sectional area function extraction device of the present invention is an extraction device that extracts a vocal tract cross-sectional area function from a digitized speech signal,
A frame cutout unit that cuts out digitized audio signals in units of frames;
A band dividing unit that divides each audio signal cut out by the frame cutout unit into a low-frequency component and a high-frequency component;
A flattening processing unit for flattening frequency characteristics of each component divided by the band dividing unit;
An autocorrelation function converting unit that converts each component processed by the flattening processing unit into an autocorrelation function;
An LSP parameter extraction unit that extracts an LSP parameter from the autocorrelation function of each component converted by the autocorrelation function conversion unit;
An LSP parameter combining unit that combines the LSP parameters of each component extracted by the LSP parameter extracting unit;
A PARCOR coefficient conversion unit for converting the LSP parameter combined by the LSP parameter combination unit into a PARCOR coefficient;
And a vocal tract cross-sectional area function conversion unit that converts the PARCOR coefficient converted by the PARCOR coefficient conversion unit into a vocal tract cross-sectional area function.

An extraction method for extracting a vocal tract cross-sectional area function of the present invention is an extraction method for extracting a vocal tract cross-sectional area function from a digitized speech signal,
Cutting the digitized audio signal into frames;
Dividing each extracted audio signal into a low frequency component and a high frequency component;
Flattening frequency characteristics of each divided component;
Converting each flattened component into an autocorrelation function;
Extracting LSP parameters from the autocorrelation function of each transformed component;
Combining the extracted LSP parameters of each component;
Converting the combined LSP parameters to PARCOR coefficients;
Converting the converted PARCOR coefficient into a vocal tract cross-sectional area function.

  In this vocal tract cross-sectional area function extraction apparatus and extraction method, a speech signal is divided into a low-frequency component and a high-frequency component, and the frequency characteristics of each component are flattened. For this reason, the sound source and the radiation characteristic can be satisfactorily removed from the low-frequency component and the high-frequency component. Thereafter, each flattened component is converted into an autocorrelation function, and LSP parameters are extracted from the autocorrelation function of each component. Since the extracted LSP parameters of each component are combined and converted from the combined LSP parameters into PARCOR coefficients, the vocal tract cross-sectional area function can be extracted without being affected by the sound source and the radiation characteristics.

  Therefore, the vocal tract cross-sectional area function extraction apparatus and extraction method of the present invention can satisfactorily control voice quality even with a high sampling frequency audio signal.

  In the flattening processing unit, an adaptive inverse filter method can be used to flatten the frequency characteristic of the low frequency component. In the flattening processing unit, a method based on cepstrum processing can be used to flatten the frequency characteristics of the high frequency component. In this case, the sound source and the radiation characteristic can be appropriately removed from each component.

  A PARCOR coefficient with a minus sign is called a reflection coefficient. For this reason, in the present invention, the PARCOR coefficient is a concept including a reflection coefficient.

It is a block diagram which shows the extraction apparatus of the vocal tract cross-sectional area function of an Example. It is a step figure which shows the extraction method of the vocal tract cross-sectional area function of an Example. Regarding the vowel “A”, FIG. (A) is a graph showing the vocal tract cross-sectional area function, and FIG. (B) is a graph showing the LSP and the spectral envelope. Regarding the vowel “I”, FIG. (A) is a graph showing the vocal tract cross-sectional area function, and FIG. (B) is a graph showing the LSP and the spectral envelope. Regarding the vowel “U”, FIG. (A) is a graph showing the vocal tract cross-sectional area function, and FIG. (B) is a graph showing the LSP and the spectral envelope. Regarding the vowel “e”, FIG. (A) is a graph showing the vocal tract cross-sectional area function, and FIG. (B) is a graph showing the LSP and the spectral envelope. Regarding the vowel “o”, FIG. (A) is a graph showing the vocal tract cross-sectional area function, and FIG. (B) is a graph showing the LSP and the spectral envelope.

  DESCRIPTION OF EMBODIMENTS An embodiment embodying a vocal tract cross-sectional area function extraction device and an extraction method thereof according to the present invention will be described with reference to the drawings.

  As shown in FIG. 1, the vocal tract cross-sectional area function extraction apparatus 1 according to the embodiment includes a frame cutout unit 10, a band division unit 20, a flattening processing unit 30, an autocorrelation function conversion unit 40, an LSP parameter extraction unit 50, An LSP parameter combining unit 60, a PARCOR coefficient converting unit 70, and a vocal tract cross-sectional area function converting unit 80 are provided.

  A method for extracting the vocal tract cross-sectional area function using the vocal tract cross-sectional area function extracting device 1 will be described.

First, as shown in FIG. 2, a step 11 for cutting out a digitized audio signal into a section called a frame is executed. Thus, the audio signal cut out in units of frames is represented as x _n (n = 1,..., N).

  The frame length is often about 30 milliseconds. The frame length may be determined by pitch synchronization analysis that is determined in synchronization with the basic period of speech. Further, the next frame is usually in a state where there is an overlap with the previous frame. The time lag from the previous frame is called the frame shift length. When the frame length is fixed, the frame shift length is 4 ms to about half of the frame length. When the frame length is determined by pitch synchronization analysis, the frame shift length is synchronized with the basic period of the voice. It is decided.

Next, step 21 for dividing each audio signal cut out in frame units into a low frequency component and a high frequency component is executed in the band dividing unit 20. That is, the audio signal x _n cut out in units of frames is Fourier transformed to obtain the audio spectrum X _k (k = 0,..., N−1). Thereafter, the voice spectrum X _k is divided into bands at the frequency f _c , that is, k _c = Nf _c / f _s .

At this time, the low frequency component is set to a minimum frequency band necessary for transmitting phonological properties in the audio signal, and the high frequency component is set to a high frequency band required for synthesizing a high quality voice. For example, the low frequency band is 0 to F _s / 8 (Hz), and the high frequency band is F _s / 8 to F _s / 2 (Hz) (where F _s is the sampling frequency). The spectrum of the divided low frequency band is expressed by Expression (1), and the spectrum of the high frequency band is expressed by Expression (2).

  Next, steps 31 and 32 for flattening the frequency characteristics of the divided components are executed in the flattening processing unit 30. That is, the spectrum in the low frequency band represented by the expression (1) is flattened by the adaptive inverse filter process which is the conventional method (step 31). A function representing this adaptive inverse filter processing is represented by Equation (3). Further, after moving the spectrum of the high frequency band represented by the expression (2) to the low band, the spectrum is flattened by removing the 0th order and the 1st order of the cepstrum (step 32). A function representing a process for removing the 0th order and the 1st order of the cepstrum is represented by Expression (4).

Next, steps 41 and 42 for converting each flattened component into an autocorrelation function are executed in the autocorrelation function converter 40. That is, the low frequency band component processed by the flattening processing unit 30 is converted into an autocorrelation function represented by the equation (5) (step 41). Further, the high frequency band component processed by the flattening processing unit 30 is converted into an autocorrelation function expressed by the equation (6) (step 42). F ⁻¹ represented in the equations (5) and (6) is an inverse Fourier transform. For N _L and N _H , those satisfying 2k _c ≦ N _L and N−2k _c ≦ N _H are selected. Although there is a section having the same value at the center of the spectrum, it is only necessary that the spectrum changes smoothly (it is not desirable to have a sharp valley in the spectrum). For example, processing such as using a smooth function is performed. May be.

Next, steps 51 and 52 for extracting LSP parameters from the converted autocorrelation function of each component are executed in the LSP parameter extraction unit 50. That is, LSP parameters are extracted from the autocorrelation functions expressed in Equation (5) and Equation (6) by linear prediction analysis and LSP analysis. The LSP parameter in the low frequency band is expressed by Expression (7), and the LSP parameter in the high frequency band is expressed by Expression (8). p _L is the analysis order in the low frequency band, and p _H is the analysis order in the high frequency band.

Next, step 61 for combining the extracted LSP parameters of each component is executed in the LSP parameter combining unit 60. That is, the LSP parameters of the respective components represented by the equations (7) and (8) are combined as shown in the equation (9) to obtain the LSP parameter ω _m in the original frequency band. Here, a technique for preventing the extreme attenuation of the spectral envelope is added by adding one of the LSP parameters serving as the midpoints of the adjacent LSP parameters between the LSP parameters of the respective components.

Here, omega _c, the normalized angular frequency of a frequency band division is calculated by _{ω c = 2πf c / f s} .

Next, a step 71 for converting the combined LSP parameter to the PARCOR coefficient is executed in the PARCOR coefficient conversion unit 70. That is, the combined LSP parameter ω _m is converted to the PARCOR coefficient κ _m through the linear prediction coefficient.

Next, a step of converting the converted PARCOR coefficient into a vocal tract cross-sectional area function is executed in the 81 vocal tract cross-sectional area function converting unit 80. That is, the PARCOR coefficient κ _m (m = 1,..., P) is converted into the vocal tract cross-sectional area function A _m (m = 1,..., P) by the equation (10).

Here, _Ag is the area of the glottal division, and it is usual to use a measurement result or a value of 1 cm, for example.

  As described above, the vocal tract cross-sectional area function extraction device and the vocal tract cross-sectional area function extracted by the conventional method using only the adaptive inverse filter method and the vocal tract cross-sectional area function extracted by the extraction method of the vocal tract cross-section function of the present embodiment described above. Are plotted in FIG. 3A to FIG. 7A. Corresponding to each, the spectral envelopes obtained by the LSP and the LSP are shown in FIGS. 3 relates to the vowel “A”, FIG. 4 relates to the vowel “I”, FIG. 5 relates to the vowel “U”, FIG. 6 relates to the vowel “E”, and FIG. 7 relates to the vowel “O”.

  3A to 7A, the horizontal axis indicates the distance from the lips, and the vertical axis indicates the logarithm of the vocal tract cross-sectional area function at that point. The base of the logarithm is 2. Moreover, the plot X of each figure (A) is a vocal tract cross-sectional area function extracted by the extraction apparatus and the extraction method of this embodiment, and the sampling frequency is 44.1 kHz. Also, plots Y and Z in each figure (A) are vocal tract cross-sectional area functions extracted by a conventional method using only the adaptive inverse filter method, and plot Y is a sampling frequency of 11.025 kHz. Plot Z shows a sampling frequency of 44.1 kHz. As shown in each figure (A), while the shape of the plot Z is greatly different from the shape of the plot Y, the plot X shows a shape close to the plot Y and a more precise shape is obtained. Has been obtained.

  In FIG. 3B to FIG. 8B, the horizontal axis indicates the frequency, and the vertical axis indicates the amplitude spectrum. As shown in each figure (B), it can be seen that the spectral envelope does not drop even in the high frequency region.

  Therefore, the vocal tract cross-sectional area function extraction device and the extraction method thereof according to the embodiment can satisfactorily control the voice quality even with a high sampling frequency audio signal.

The present invention is not limited to the embodiments described with reference to the above description and drawings. For example, the following embodiments are also included in the technical scope of the present invention.
(1) In the embodiment, a series of processing is performed in the frequency domain, but may be performed in the time domain.
(2) In this embodiment, when combining the LSP parameters of each component, an LSP parameter that is the midpoint of the adjacent LSP parameters is added between the two LSP parameters, thereby reducing the extreme spectral envelope attenuation. Although it is prevented, it is not necessary to add the LSP parameter as the midpoint.

  The present invention can be used for text-to-speech synthesis and voice quality conversion systems.

DESCRIPTION OF SYMBOLS S ... Voice signal 1 ... Extraction apparatus 10 ... Frame extraction part 20 ... Band division part 30 ... Flattening process part 40 ... Autocorrelation function conversion part 50 ... LSP parameter extraction part 60 ... LSP parameter combination part 70 ... PARCOR coefficient conversion part 80 ... Vocal tract cross-sectional area function converter

Claims (2)

An extraction device for extracting a vocal tract cross-sectional area function from a digitized speech signal,
A frame cutout unit that cuts out digitized audio signals in units of frames;
A band dividing unit that divides each audio signal cut out by the frame cutout unit into a low-frequency component and a high-frequency component;
A flattening processing unit for flattening frequency characteristics of each component divided by the band dividing unit;
An autocorrelation function converting unit that converts each component processed by the flattening processing unit into an autocorrelation function;
An LSP parameter extraction unit that extracts an LSP parameter from the autocorrelation function of each component converted by the autocorrelation function conversion unit;
An LSP parameter combining unit that combines the LSP parameters of each component extracted by the LSP parameter extracting unit;
A PARCOR coefficient conversion unit for converting the LSP parameter combined by the LSP parameter combination unit into a PARCOR coefficient;
An apparatus for extracting a vocal tract cross-sectional area function, comprising: a vocal tract cross-sectional area function converting unit that converts the PARCOR coefficient converted by the PARCOR coefficient converting unit into a vocal tract cross-sectional area function. An extraction method for extracting a vocal tract cross-sectional area function from a digitized speech signal,
Cutting the digitized audio signal into frames;
Dividing each extracted audio signal into a low frequency component and a high frequency component;
Flattening frequency characteristics of each divided component;
Converting each flattened component into an autocorrelation function;
Extracting LSP parameters from the autocorrelation function of each transformed component;
Combining the extracted LSP parameters of each component;
Converting the combined LSP parameters to PARCOR coefficients;
Converting the converted PARCOR coefficient into a vocal tract cross-sectional area function, and extracting the vocal tract cross-sectional area function.