JP5325130B2 - LPC analysis device, LPC analysis method, speech analysis / synthesis device, speech analysis / synthesis method, and program - Google Patents

Description

  The present invention relates to a speech analysis / synthesis device and a speech analysis / synthesis method for extracting a sound source signal and a vocal tract spectrum from a speech signal and synthesizing a speech signal by driving a vocal tract filter with the sound source signal. The present invention relates to an LPC analysis apparatus and an LPC analysis method for extracting a road spectrum.

  Up to now, in order to improve the performance of speech synthesis / coding / recognition technology, it is important to decompose speech signals into sound source signals and vocal tract spectrum efficiently and accurately based on human speech generation mechanism. It has been. For this decomposition, linear prediction (LPC) analysis is widely used. However, since white noise is assumed as a sound source signal, the obtained vocal tract spectrum is not a little affected by the fundamental frequency (F0). was there. In particular, since the fundamental frequency is high in female voice and the above assumption is not satisfied, the fundamental frequency of the sound source signal and its harmonics are included in the vocal tract spectrum estimated by LPC analysis, and an accurate vocal tract spectrum is obtained. There was a problem that was difficult.

  By the way, in order to avoid the problem of the fundamental frequency of the sound source signal in the LPC analysis, a sample selection linear prediction method has been proposed (for example, see Non-Patent Document 1). This is because the audio signal at the time when the LPC residual signal obtained by passing the audio signal through the LPC inverse filter becomes large does not satisfy the assumption that the white noise is a sound source signal. This is a method of performing LPC analysis. As a result, it is expected to obtain a vocal tract spectrum with little influence of the fundamental frequency, but since the phase characteristics included in the audio signal are not taken into consideration, an error occurs in the selection of the audio signal to be removed, and the result is obtained. There is a problem that the accuracy of the vocal tract spectrum is insufficient. Moreover, since the audio signal is removed, there is a possibility that the number of data used for analysis is insufficient.

  On the other hand, a method called AR-HMM has been proposed in order to solve the problem of the phase characteristics included in the audio signal (see, for example, Non-Patent Document 2). In this method, it is possible to perform robust LPC analysis at the fundamental frequency by assuming sound source modeling by HMM. However, since the sound source model is complicated including phase characteristics, it takes time to calculate, Has a problem that it is difficult to find a stable solution easily.

Yoshiaki Miyoshi, Kazuharu Yamato, Masuzou Yanagida, Kakusho, “Analysis of high-pitch speech using two-stage sample selection linear prediction method”, IEICE Transactions, Vol.J70-A, No.8, pp.1146- 1156, 1987. Minoru Sasou and Seino Tanaka, “Modeling of sound source by HMM and extraction of vocal tract characteristics robust to high fundamental frequencies”, IEICE Transactions, Vol.J84-D-II, No.9, pp.1960-1969 , 2001.

  In view of the above situation, an object of the present invention is to efficiently obtain an accurate and stable vocal tract spectrum that is not affected by a fundamental frequency from an audio signal.

  According to this invention, the LPC analyzer receives the phase equalized audio signal and the pitch mark time group as input, and the sound source signal has a single pulse of amplitude G at each pitch mark time of the pitch mark time group, and the pitch mark time. The time other than is made up of white noise, and the LPC coefficient and the amplitude G are obtained so that the error between the audio signal obtained from the LPC coefficient and the sound source signal and the phase equalized audio signal is minimized.

  A speech analysis / synthesis apparatus according to the present invention includes a speech section detection unit that detects a speech section of an input speech signal, a fundamental frequency analysis unit that estimates a fundamental frequency from the speech signal for the speech section, and the fundamental frequency The speech signal is cut out with the window length determined based on the LPC analysis, the LPC analysis is performed, and the speech signal is passed through the LPC inverse filter to obtain the LPC residual signal, and the fundamental period obtained from the fundamental frequency A pitch waveform analysis unit that generates a corresponding pitch waveform and extracts a pitch mark time group using the pitch waveform and the LPC residual signal, and uses the pitch mark time group and the LPC residual signal. Applying a phase equalization filter to the audio signal to generate a phase equalized audio signal; Each pitch mark time of the time group has a single pulse of amplitude G, and the time other than the pitch mark time is composed of white noise, the audio signal obtained from the LPC coefficient and the sound source signal, and the phase equalization Using the second LPC analysis unit for obtaining the LPC coefficient and the amplitude G so as to minimize the error from the audio signal, the pitch mark time group, the phase equalized audio signal, and the LPC coefficient, a pulse gain and a multi-value are obtained. A multi-pulse sound source model generation unit that obtains a pulse sound source model, a white noise gain generation unit that calculates a white noise gain using a PARCOR coefficient k and an autocorrelation function R obtained in the LPC analysis in the second LPC analysis unit; The white frequency multiplied by the white noise gain is used outside the voice interval, and the fundamental frequency and the pulse are used in the voice interval. A voice signal is synthesized by convolution calculation of a sound source signal using a multi-pulse calculated from the gain and the multi-pulse sound source model or a single pulse train calculated from the fundamental frequency and the pulse gain and the LPC coefficient. It consists of a speech synthesizer.

  According to the present invention, an accurate and stable vocal tract spectrum that is not affected by the fundamental frequency can be obtained efficiently.

The block diagram which shows the function structure of one Example of the speech analysis synthesis apparatus by this invention. The flowchart (the 1) which shows the flow of a process in the speech analysis and synthesis apparatus shown in FIG. The flowchart (the 2) which shows the flow of a process in the speech analysis and synthesis apparatus shown in FIG. The flowchart (the 3) which shows the flow of a process in the speech analysis and synthesis apparatus shown in FIG. The graph which shows the analysis result of a vocal tract spectrum. The figure which shows the analysis result of a vocal tract spectrum series. The figure which shows the example of a waveform of each process in a speech analysis synthesis process.

  Embodiments of the present invention will be described with reference to the drawings.

  FIG. 1 shows a functional configuration of an embodiment of a speech analysis / synthesis apparatus according to the present invention. In this example, a speech analysis / synthesis apparatus 10 includes a speech section detection unit 11, a fundamental frequency analysis unit 12, and a first LPC analysis unit 13. And a pitch mark analysis unit 14, a phase equalized speech generation unit 15, a second LPC analysis unit 16, a multi-pulse sound source model generation unit 17, a white noise gain generation unit 18, and a speech synthesis unit 19.

2 to 4 show the flow of processing in the speech analysis / synthesis apparatus 10 shown in FIG. 1, and the function of each unit and the flow of processing will be described below with reference to FIGS.
<Audio section detection unit>
First, the voice section detection unit 11 detects a voice section based on the power threshold processing of the voice signal (original voice) (step S1).

<Basic frequency analysis section>
Next, the fundamental frequency analysis unit 12 estimates a fundamental frequency from the speech signal using a pitch extraction algorithm for the obtained speech section. For example, in the present embodiment, the fundamental frequency is obtained based on the instantaneous frequency amplitude spectrum with an analysis window length (analysis interval) of 30 ms and an analysis shift length of 4 ms (step S2). For analysis of the fundamental frequency, for example, a technique based on the instantaneous frequency amplitude spectrum described in the following document A can be used.
Reference A: Arifianto, D., Tanaka, T., Masuko, T., and Kobayashi, T., “Robust F0 estimation of speech signal using harmonicity measure based on instantaneous frequency”, IEICE Trans. Information and Systems, Vol. E87 -D, No.12, pp.2812-2820,2004.

<First LPC analysis unit>
In order to obtain an LPC residual signal used for phase equalization processing, the first LPC analysis unit 13 uses an analysis shift length of 4 ms and sets the voice signal to 2.5 times the fundamental period (basic period = 1 ÷ basic frequency) as the window length. The extracted Blackman window is used to perform LPC analysis by the autocorrelation method (step S3). Then, the LPC residual signal is obtained by passing the audio signal through the LPC inverse filter (step S4).

  In the phase equalization process, the flatness of the spectrum of the LPC residual signal is a condition for preventing the spectrum of the speech signal from being changed. Therefore, in this embodiment, for the purpose of flattening the spectrum of the LPC residual signal, The analysis order of LPC is set higher (for example, 50th for male voice and 40th for female voice). In order to avoid the influence of the fundamental frequency, a lag window (100 Hz) is used.

Furthermore, since there is a problem that the analysis of the power spectrum using the window function depends on the analysis time, a TANDEM window as described in the following document B is used for the purpose of smoothing the vocal tract spectrum in the time direction.
Reference B: Masamasa Mori, Toru Takahashi, Hidenori Kawahara, Toshio Irino, “Speech Analysis Using Power Spectrum Estimation Method for Periodic Signals Independent of Analysis Time”, IEICE Transactions, Vol. J92-A, No.3, pp.163-171, 2009.

  This is a method of estimating a power spectrum that does not depend on the analysis time by adding the power spectrum of the analysis frame and the analysis frame shifted by half of the basic period and dividing by two. According to Wiener Hinting's theorem, the inverse Fourier transform of the power spectrum is an autocorrelation function, so when using it in LPC analysis by the autocorrelation method, add the autocorrelation function of the analysis frame and the analysis frame shifted by half the fundamental period. LPC coefficients can be obtained by dividing the obtained autocorrelation matrix by the Durbin algorithm.

Here, the basic period contained in the analysis frame if the constant T _0, the shift amount required for the calculation of the TANDEM window that may be a T _0/2 are shown in the literature B. However, in an actual audio signal, the fundamental period in the analysis frame is not constant, such as T ₀ , T ₁ , and in this embodiment, the frequency space is used as a shift amount considering the fluctuation of the fundamental period in the analysis frame. Weighted average at

Is used. Here, w is a Gaussian weight.

<Pitch mark analysis unit>
The pitch mark analysis unit 14 generates a pulse series signal (pitch waveform) corresponding to the fundamental period obtained from the fundamental frequency within the speech section in order to obtain a pitch mark (pitch mark time group) used for phase equalization processing. (Step S5). At frame number t and time k, the cross-correlation is performed for each frame t between the absolute value of the pitch waveform ex (t, k) and the absolute value of the LPC residual signal e (t, k) within the speech section. Function r (t, j) = Σ _k | e (t, k) | × | ex (t, k + j) |
Was calculated, the Σ _{t r (t,} j) is series of j such that maximum, determined using a dynamic programming method to obtain a candidate of the pitch mark time group. Then, a time at which the absolute value of the LPC residual signal is maximized is searched in the vicinity of the obtained pitch mark time.

  Furthermore, the pitch mark time at which the absolute value of the residual is the maximum is selected from the obtained pitch mark time group, and the autocorrelation function (modified self) of the LPC residual signal between adjacent pitch mark times is selected. The time when the correlation function is maximized is sequentially searched and extracted as a final pitch mark time group (step S6). Here, a difference may be taken between the obtained adjacent pitch mark times and used as a new basic period.

<Phase equalized speech generator>
In order to obtain a phase-equalized audio signal, the phase-equalized audio generation unit 15 uses the pitch mark (pitch mark time group) and the LPC residual signal, and inverts the value of the LPC residual signal around the pitch mark time. Then, a phase equalization filter having a normalized value as a coefficient is obtained, and this is applied to an audio signal (hereinafter referred to as the original audio) to obtain a phase equalized audio signal (step S7).

  Here, after the phase equalization filter coefficient obtained at each pitch mark time is smoothed temporally using, for example, a first-order low-pass filter, the phase equalization filter is applied to the original sound. The number of taps of the phase equalization filter is the same as the length of the fundamental period. The deformation of the original speech spectrum due to the phase equalization process is slight. In addition, since human hearing is relatively insensitive to the short-time phase characteristics of audio signals, the difference in audibility between the original audio and the phase-equalized audio is slight. This processing is based on, for example, the method described in Japanese Patent No. 2061816 (hereinafter referred to as Patent Document 1), and by concentrating the energy of the LPC residual signal over time, Makes it possible to approximate the included phase to zero.

<Second LPC analysis unit>
The second LPC analysis unit 16 performs LPC analysis on the phase equalized speech signal acquired by the phase equalized speech generation unit 15.
Here, it can be assumed that the phase-equalized audio signal is generated by passing a sound source signal of a single pulse train and white noise through a vocal tract filter. That is, the amplitude of the pulse in the group of I + 1 pitch marks included in the analysis frame (t ₀ , t ₁ ,..., T _I is the pitch mark time) is G, and other times t are the same as in the conventional LPC analysis. Assuming white noise as the sound source,

This results in the problem of obtaining the LPC coefficient a and the amplitude G of the pulse from the phase-equalized audio signal. Here, r is a weight, e is an LPC residual signal, s is a phase equalized speech signal, and p is an LPC analysis order.

The phase-equalized audio signal is cut out by a Blackman window with an analysis shift length of 4 ms and a window length of 2.5 times the basic period. If r = 0, this is equivalent to the sample selection linear prediction method in which s (t ₀ )... s (t _I ) of the phase-equalized audio signal is removed. If G = 0 and r = 1, This is equivalent to LPC analysis assuming white noise.

  In this embodiment, by setting G ≠ 0 and r = 1, the problem of lack of data due to the removal of phase-equalized speech, which was a drawback of the sample selection linear prediction method, is avoided, and further, LPC analysis and sound source The characteristic is that the signal parameter G can be optimized in the same framework. The LPC coefficient that minimizes the evaluation formula can be obtained by solving the following simultaneous equations (provided that r = 1).

This simultaneous equation can be efficiently solved using the Levinson algorithm. Here, the autocorrelation function R of the phase equalized speech signal is

The amplitude G of the pulse is obtained from the PARCOR coefficient k and the autocorrelation function R obtained when solving the Levinson algorithm.

It becomes. Since the above equation (3) is an extension of the autocorrelation method, it is possible to apply a TANDEM window for the purpose of temporal smoothing of the vocal tract spectrum. In this case, as the value of the cross-correlation function including G on the right side, a value obtained by adding the cross-correlation function of the analysis frame and the analysis frame shifted by the expression (1) and dividing by 2 may be used.

  Assuming that the initial value of G is calculated from the above equation (5) using the PARCOR coefficient k and autocorrelation function R obtained when solving the Durbin algorithm by calculating the phase equalized speech signal, the LPC coefficient is determined (step S9) and G determination (step S8) is repeated (for example, 5 times). This repetition is expected to improve the estimation accuracy of the obtained parameters. In this embodiment, it can be assumed that the phase-equalized audio signal is generated from a relatively simple sound source, so that the problem to be solved is simplified. This leads to the result that iterative calculation does not take time and a stable solution can be easily obtained. The analysis order of the LPC analysis here may be different from the order for the phase equalization process. In addition, since the influence of the fundamental frequency is reduced by this analysis, the lag window need not be used.

  Note that the LPC coefficient and white noise gain outside the speech section are obtained using a fixed window length of 15 ms and a fixed frame shift length of 4 ms. This is equivalent to setting I = 0 in the LPC analysis method of this embodiment.

<Multipulse source model generator>
The multi-pulse sound source model generation unit 17 sets the amplitude of the pulse (pulse gain) and the parameters (FIR filter) of the multi-pulse sound source model so that the auditory weighted error with the phase-equalized speech signal is minimized within the speech section. The coefficient v _k ) is _obtained (step S10). Here, the transfer characteristic of the FIR filter (6 taps) is expressed as follows, as in Patent Document 1.

For the calculation of the phase equalization pulse sound source, each pitch mark time position is set as the analysis start time, and the analysis window length is determined as one basic period. In this embodiment, pitch synchronization analysis is used for the analysis, but since a 4 ms frame shift is used for the synthesis, there is a problem that the pitch mark time position is different from the start time of the fixed-length frame. Therefore, in this embodiment, the parameters in each frame are obtained by linear interpolation. In this embodiment, since a smooth LPC spectrum is obtained in both time and frequency, the parameters of the sound source model obtained by the above calculation are expected to change smoothly in time.

<White noise gain generator>
The white noise gain generation unit 18 calculates the white noise gain using the PARCOR coefficient k and the autocorrelation function R obtained when the second LPC analysis unit 16 solves the Levinson algorithm (step S11).

<Speech synthesis unit>
The speech synthesizer 19 performs speech synthesis by convolving the LPC coefficient and the sound source signal (step S13). As a sound source signal, a signal obtained by multiplying white noise by a white noise gain is used outside the speech section. In the speech section, a multipulse calculated from the fundamental frequency, pulse gain, and multipulse sound source model, or a single pulse train calculated from the fundamental frequency and pulse gain is used (step S12).

[Modification]
Bandpass filter of 0-500, 500-1000, 1000-2000, 2000-3000, 3000-4000, 4000-5000, 5000-6000, 6000-7000, 7000-8000 Hz The autocorrelation function of each passed signal may be calculated and obtained as the voiced intensity. In this calculation, similarly to the above, pitch synchronization analysis is performed with each pitch mark time position as the analysis start time. When synthesizing speech, based on the voiced intensity, a voice source is created with a band greater than a certain threshold as a voiced band, a smaller band as a voiceless band, a multipulse or single pulse train in the voiced band, and white noise mixed in the voiceless band. Then, it may be convolved with the LPC coefficient.

  Further, after converting the parameters obtained by the voice analysis, the voice may be synthesized. For example, the fundamental frequency value can be halved or doubled, the formant frequency obtained by the LPC analysis can be arbitrarily shifted, or the time axis can be doubled or halved.

[Experimental example]
FIG. 5 shows the LPC spectrum of “rise” produced by a female English native speaker. In this experiment, the LPC analysis order for phase equalization processing was 40th, and the analysis order for obtaining an LPC spectrum was 20th. The audio sampling rate is 16 kHz. The thin line in FIG. 5 is obtained by the conventional LPC analysis method using a lug window, and the thick line is obtained by the LPC analysis method (proposed method) according to the present invention.

  In the conventional method, the fundamental frequency is picked up as a vocal tract spectrum. However, it can be seen that the proposed method can accurately extract the first formant frequency (F1) without picking up the fundamental frequency. It can also be seen that the amplitude of F2 has recovered with the improvement of F1.

  FIGS. 6A and 6B are LPC spectra of “skills” uttered by a male speaker, where FIG. 6A shows no TANDEM window (conventional method) and FIG. 6B shows a TANDEM window (proposed method). The LPC analysis order for the phase equalization process and the LPC spectrum was 50th. (A) shows a temporally discontinuous spectrum (that is, vertical stripes are seen), but (B) shows that a temporally continuous spectrum is obtained.

  FIG. 7 shows a partial waveform example of the utterance material “skill”. 7A is an audio signal sampled at 16 kHz, FIG. 7B is an LPC residual signal obtained by performing LPC analysis and passing the audio signal through an LPC inverse filter, and FIG. 7C is an LPC signal. The phase equalization residual signal obtained by applying the phase equalization filter based on the residual signal to the LPC residual signal. FIG. 7D shows the phase equalization sound obtained by applying the phase equalization filter to the audio signal. FIG. 7E shows a synthesized audio signal.

  In FIG. 7B, the effect of the phase characteristic is also seen in the LPC residual signal, but FIG. 7C in which the phase equalization processing is performed can be regarded as a mixture of a single pulse train and white noise. Furthermore, although FIG. 7D has a waveform different from that of FIG. 7A, it has been confirmed that almost no difference in audibility is felt. FIG. 7 (E) shows synthesized speech, but has almost the same waveform as FIG. 7 (D), indicating the effectiveness of the proposed method. In addition, it was confirmed that there was almost no difference in audibility between synthesized speech and original speech by the proposed method.

  The speech analysis / synthesis apparatus and speech analysis / synthesis method described above can be realized by a computer and a program installed in the computer. The computer functions as a speech analysis / synthesis device by executing the installed program.

  Note that the second LPC analysis unit in the speech analysis / synthesis apparatus 10 shown in FIG. 1 may be handled as a single LPC analysis apparatus.

Claims (7)

With the phase equalized audio signal and pitch mark time group as inputs,
The sound source signal has a single pulse of amplitude G at each pitch mark time of the pitch mark time group, and the time other than the pitch mark time is composed of white noise,
The LPC coefficient and the amplitude G are obtained using the PARCOR coefficient k and the autocorrelation function R so that the error between the audio signal obtained from the LPC coefficient and the sound source signal and the phase equalized audio signal is minimized. An LPC analyzer characterized by being configured. With the phase equalized audio signal and pitch mark time group as inputs,
The sound source signal has a single pulse of amplitude G at each pitch mark time of the pitch mark time group, and the time other than the pitch mark time is composed of white noise,
The LPC coefficient and the amplitude G are obtained using the PARCOR coefficient k and the autocorrelation function R so that the error between the audio signal obtained from the LPC coefficient and the sound source signal and the phase equalized audio signal is minimized. An LPC analysis method characterized by the above. A voice section detector for detecting a voice section of the input voice signal;
A fundamental frequency analyzer for estimating a fundamental frequency from the speech signal for the speech section;
A first LPC analysis unit that cuts out the speech signal with a window length determined based on the fundamental frequency, performs LPC analysis, and obtains an LPC residual signal by passing the speech signal through an LPC inverse filter;
A pitch mark analysis unit that generates a pitch waveform according to a fundamental period obtained from the fundamental frequency and extracts a pitch mark time group using the pitch waveform and the LPC residual signal;
A phase-equalized sound generation unit that generates a phase-equalized sound signal by applying a phase equalization filter obtained using the pitch mark time group and the LPC residual signal to the sound signal;
The sound source signal has a single pulse of amplitude G at each pitch mark time of the pitch mark time group, and a time other than the pitch mark time is composed of white noise, and a sound signal obtained from the LPC coefficient and the sound source signal And a second LPC analysis unit for obtaining the LPC coefficient and the amplitude G using the PARCOR coefficient k and the autocorrelation function R so that an error from the phase equalized speech signal is minimized,
Using the pitch mark time group, the phase-equalized audio signal, and the LPC coefficient, a multi-pulse sound source model generating unit for obtaining a pulse gain and a multi-pulse sound source model;
A white noise gain generation unit that calculates a white noise gain using the PARCOR coefficient k and the autocorrelation function R obtained in the LPC analysis in the second LPC analysis unit;
Other than the speech section, white noise multiplied by the white noise gain is used. In the speech section, multipulses calculated from the fundamental frequency, the pulse gain, and the multipulse sound source model, or the fundamental frequency and the pulse gain. A speech synthesizer that synthesizes a speech signal by performing a convolution operation on a sound source signal using a single pulse train calculated from the LPC coefficient,
A speech analysis / synthesis apparatus characterized by comprising: The speech analysis / synthesis device according to claim 3,
The second LPC analysis unit calculates an initial value of the amplitude G using a PARCOR coefficient k and an autocorrelation function R obtained at the time of LPC analysis in the first LPC analysis unit. . A voice segment detection process for detecting a voice segment of the input voice signal;
A fundamental frequency analysis process for estimating a fundamental frequency from the speech signal for the speech interval;
A first LPC analysis process in which the speech signal is cut out with a window length determined based on the fundamental frequency to perform LPC analysis, and an LPC residual signal is obtained by passing the speech signal through an LPC inverse filter;
A pitch mark analysis process for generating a pitch waveform corresponding to a fundamental period obtained from the fundamental frequency and extracting a pitch mark time group using the pitch waveform and the LPC residual signal;
A phase-equalized speech generation process for generating a phase-equalized speech signal by applying a phase-equalization filter obtained using the pitch mark time group and the LPC residual signal to the speech signal;
The sound source signal has a single pulse of amplitude G at each pitch mark time of the pitch mark time group, and a time other than the pitch mark time is composed of white noise, and a sound signal obtained from the LPC coefficient and the sound source signal And a second LPC analysis process for obtaining the LPC coefficient and the amplitude G using the PARCOR coefficient k and the autocorrelation function R so that an error from the phase equalized speech signal is minimized,
Using the pitch mark time group, the phase-equalized audio signal, and the LPC coefficient, a multi-pulse sound source model generation process for obtaining a pulse gain and a multi-pulse sound source model;
A white noise gain generation step of calculating a white noise gain using a PARCOR coefficient k and an autocorrelation function R obtained in the LPC analysis in the second LPC analysis step;
Other than the speech section, white noise multiplied by the white noise gain is used. In the speech section, multipulses calculated from the fundamental frequency, the pulse gain, and the multipulse sound source model, or the fundamental frequency and the pulse gain. A speech synthesis process for synthesizing a speech signal by performing a convolution operation on a sound source signal using a single pulse train calculated from the LPC coefficient;
A speech analysis and synthesis method characterized by comprising: The speech analysis and synthesis method according to claim 5,
In the second LPC analysis process, an initial value of the amplitude G is calculated using a PARCOR coefficient k and an autocorrelation function R obtained in the LPC analysis in the first LPC analysis process. .   A program for causing a computer to function as the speech synthesis analysis apparatus according to claim 3.