JP2898637B2 - Audio signal analysis method - Google Patents

Description

Description: Object of the Invention (Industrial application field) The present invention relates to a speech signal analysis method of an encoding system for compressing a speech signal with high efficiency, and particularly to a pole-zero model analysis method. The present invention relates to an audio signal analysis method.

(Prior Art) A method of analyzing a speech signal based on linear prediction is widely used in speech coding, an analysis / synthesis system, recognition, and the like. This analysis method approximates the envelope of the short-time spectrum of speech with an all-pole model (also called an AR model), and closely approximates the poles of the spectrum. But,
Zeros have characteristics that cannot be approximated well. Since real speech has zeros in the spectrum like consonants and nasal sounds,
It has been considered that an all-pole model is not sufficient for accurately analyzing audio signals.

Therefore, several methods have been proposed for approximating the short-time spectral envelope of speech using a pole-zero model (also referred to as an ARMA model). For example, "Fixed order of pole-zero model in speech analysis" (IEICE (A). Vol. J60-A, No. 4, PP4
23-424 (1977-4)).

In this method, a short-time spectrum is approximated by a pole-zero transfer function shown below.

(Where a _i and p are the parameters and order of the polar model, respectively.
b _i and q: parameters and order of the zero model) This pole-zero model can be said to well represent the actual speech generation process. Assuming that the contents of the above-mentioned document are configured as shown in FIG. 9, the input speech signal is first converted to an all-pole model (b _i = 0, i = 1, 2,..., q)
Analysis is performed. Specifically, the autocorrelation processing unit
According to 101, the parameters a _i (i = 1, 2,...,
p) is required. Next, the audio signal is passed through an all-pole inverse filter 102 to become a residual signal from which the poles of the spectrum have been removed.

Next, the power spectrum of the residual signal is obtained by the power spectrum unit 10 and the reciprocal is obtained by the reciprocal processing unit 104. As a result, the zeros remaining in the spectrum of the residual signal are replaced with poles. Therefore, it is possible to apply the analysis of the all-pole model autocorrelation coefficients inverse Fourier transform is obtained is subjected to inverse power spectrum by inverse Fourier transform unit 105, zero model parameters b _{i (i} = 1 , 2, ..., q).

However, when applying this method to real speech,
There is a big problem. This is because the fine structure of the spectrum due to the pitch causes a large error in estimating the zero model parameters. To explain in detail, the spectrum of a male audio signal input by the μ-PCM codec having a frequency of 8 kHz in FIG. 10, as can be seen from FIG. 11, which shows an example of the spectrum of the residual signal after polar prediction in FIG. The fine structure based on the pitch of the sound source appears in the spectrum of the residual signal. The deep valleys of this microstructure act like zeros or emphasize true zeros. For example, this is shown at point A in the spectrum of the residual signal in FIG. Such a zero will be referred to as a false zero or an enhanced zero. Therefore, when analyzed by the above-described conventional method, the frequency characteristic of the obtained pole-zero model is far apart from the spectrum envelope of the audio signal as shown in FIG.

As described above, in the conventional method of analyzing the pole-zero model, there is a problem that the estimation of the zero point is erroneous due to the fine structure of the spectrum due to the pitch, and the spectrum approximation is worse than that of the all-pole model. There is. As a method of solving this problem, a method of smoothing the power spectrum of the residual signal by two-point averaging can be easily considered. However, the pitch period varies greatly depending on individuals, particularly on gender, and even on the voice of the same person. Will vary. Therefore, in the method of simply averaging the power spectrum, the smoothing of the power spectrum is not always performed satisfactorily.

(Problems to be Solved by the Invention) As described above, in the conventional voice analysis method using the pole-zero model, zero points cannot be accurately extracted due to the influence of pitch, and the spectrum approximation of the obtained model is deteriorated. There is.

The present invention has been made in view of such a problem, and an object thereof is to provide a parameter of a pole-zero model that can always accurately extract zeros without being affected by pitch and without depending on a speaker or a phoneme. An audio signal analysis method including an extraction method is provided.

[Configuration of the invention]

(Means for Solving the Problems) The present invention smoothes the power spectrum of the residual signal of the all-pole model or the reciprocal of the power spectrum in the time domain, and performs inverse Fourier transform from the reciprocal of the smoothed power spectrum. The autocorrelation coefficient is obtained, and a parameter of a zero point is extracted by applying an all-pole model analysis method to the obtained autocorrelation coefficient. However, the degree of smoothing is adaptively changed according to the value of the pitch period. Also,
When a filter is used for smoothing, the filter used has zero phase.

(Operation) First, an input signal is analyzed using an all-pole model, and parameters of the all-pole model are extracted. Next, the input signal is passed through an all-pole inverse filter to obtain a residual signal. The spectrum of the residual signal is the one with the poles removed, then the inverse of the power spectrum is taken and the zeros of the spectrum are converted to poles. At this time, the fine structure based on the pitch is smoothed by smoothing the spectrum in the time domain. The interval of the fine structure of the spectrum due to the pitch is proportional to the reciprocal of the pitch period, and varies from person to person or even from the same person due to phonemes. However, according to the present method, the degree of spectrum smoothing is adaptively changed according to the pitch cycle, so that the spectrum can always be satisfactorily smoothed regardless of the speaker or phoneme. False zeros and overemphasized zeros due to the structure can be removed. Further, in the present method, the problem that the zero point of the spectrum due to the phase characteristic of the filter shifts is prevented by setting the filter used for smoothing to zero phase. As a result, a pole-zero model that satisfactorily approximates the spectrum of speech can be obtained.

(Example) Hereinafter, an example according to the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing a pole-zero prediction analysis method according to an embodiment of the present invention. In the figure, an audio signal is input to a terminal 1 and is input to a polar parameter estimator 2. Several methods are known for estimating the polar parameters.
For example, the document "Digital Voice Processing" (Tokai University Press)
Can be used. The input signal is then input to an all-pole inverse filter 3 having the pole parameters obtained by the pole parameter estimator 2. Here, the prediction residual signal d (n) is calculated and output according to the following equation.

Here, S (n) is the input signal series, s _i is the parameter of the all-pole model, p is a prediction order.

Next, a fast Fourier transform (FFT) unit 4 and a squaring circuit unit 5
, The power spectrum of the residual signal d (n) is obtained, and the pitch analysis processing unit 6 extracts a pitch cycle and determines voiced / unvoiced. Note that a discrete Fourier transform (DFT) can be used instead of the FFT unit 4. As a pitch analysis method, for example, a modified correlation method described in the above-mentioned document “Digital Speech Processing” can be used.

The power spectrum of the residual signal obtained by the FFT unit 4 and the squaring circuit unit 5 is input to the smoothing circuit unit 7. The smoothing circuit unit 7 smoothes the power spectrum using the pitch period and the voiced / unvoiced state obtained by the pitch analysis processing unit 6 as parameters.

FIG. 3 is a block diagram showing a specific example of a smoothing circuit unit according to one embodiment of the present invention. The time constant of this circuit, that is, the number of samples T at which the impulse response becomes 1 / e
Is expressed as T = −1 / ln (α) (3) This time constant T is adaptively changed according to the value of the pitch period. If the pitch period is T _p [sample], the sampling frequency is is _[Hz] , and the order of the FFT or DFT is N, the period m [sample] of the fine structure due to the pitch appearing in the power spectrum of the residual signal is expressed by the following equation. Can be described.

Therefore, to adaptively change the time constant T according to m, Substituting equation (3) into It should be determined. Here, L is a parameter representing the number of fine structures to be smoothed. Also, the case of no speech, since T _p can not be obtained, when the pitch analysis unit 6 determines that unvoiced is set in advance appropriately determined values of T _p. Further, when the power spectrum is smoothed by the filter shown in FIG. 2, the filter has a zero phase. In order to make the phase zero, for example, the power spectrum may be filtered forward and backward, respectively, and the obtained outputs may be averaged. The power spectrum of the residual signal is represented by D (nω
₀ ), the filter output when filtering forward is D (nω ₀ ) _f , and the filter output when filtering backward is D (nω ₀ ) _b , the smoothing is described as follows.

Here, D (nω ₀ ) is a smoothed power spectrum, and N is the order of FFT or DET.

FIG. 3 shows an example of the spectrum of the smoothed residual signal. However, the spectrum was obtained by 265 point FFT.

The spectrum smoothed by the smoothing circuit described above is converted by the reciprocal circuit unit 8 into an inverse spectrum. As a result, the zeros of the residual signal spectrum are converted to poles. The inverse spectrum is subjected to inverse FFT by an inverse FFT processing unit 9, converted into an autocorrelation sequence, and input to a zero prediction parameter estimation unit 10. The zero prediction parameter estimating unit 10 obtains a zero prediction parameter from the input autocorrelation sequence using an autocorrelation method. The all-zero inverse filter 11 receives the residual signal of the all-pole inverse filter as an input and outputs a zero prediction parameter estimator 10.
The prediction is performed using the zero prediction parameter obtained by the above, and a prediction residual signal e (n) is output. e (n) is calculated according to the following equation.

Here, b _i is zero prediction parameter, Q is the order of the zero prediction.

With the above processing, the pole-zero prediction analysis of the audio signal is performed. FIG. 4 shows the frequency characteristics of the pole-zero model obtained.

The smoothing circuit shown in FIG. 1 can also be implemented by a method of detecting peaks of a power spectrum and interpolating between the detected peaks with a quadratic curve. Specifically, the coefficients of the quadratic equation passing through the three peaks are obtained, and the interpolating between the two peaks is performed using the quadratic curve. In this case, there is an effect that the amount of calculation is reduced because pitch analysis is not required.

The smoothing circuit shown in FIG. 1 can be inserted next to the reciprocal circuit, and an embodiment in this case is shown in FIG. The smoothing shown in FIGS. 1 and 5 performed in the frequency domain can be performed in the time domain. Residual signal d
The reciprocal of the power spectrum of (n) is D '(nω ₀ ),
(N = 0, 1,..., N−1) If the impulse response and the transfer function of the digital filter shown in FIG. 2 are denoted by h (n) and H (nω ₀ ), the smoothing can be expressed by the following equation. This is performed by filtering in the frequency domain.

Here, (nω ₀ ) is a smoothed power spectrum. If the inverse Fourier transforms of (nω ₀ ) and D ′ (nω ₀ ) are (n) and γ ′ (n), respectively, equation (11) is described in the time domain as Is done.

(N) = γ ′ (n) · H (nω ₀ ) (13) That is, it is equivalent to multiplying the window H (nω ₀ ). At this time, H (nω ₀ ) is called a lag window. H (n
ω ₀ ) changes adaptively according to the pitch period.

FIG. 6 shows an embodiment in which the smoothing is performed in the time domain.

Further, in the embodiments shown in FIGS. 1, 5 and 6, the conversion of the zero point to the pole is performed in the frequency domain, but this can be performed in the time domain. Pole autocorrelation sequence of the residual signal d (n) of the prediction gamma (n), placing the power spectrum is the Fourier transform and D (nω _0), D the inverse D '(nω ₀ of (nω ₀₎ ) Has the following relationship:

D (nω ₀ ) · D ′ (nω ₀ ) = 1 (14) From the nature of the Fourier transform, the above equation is expressed in the time domain as follows.

Since the autocorrelation coefficient is symmetric about γ (0), this equation (15) can be written as a matrix as follows.

This equation can be solved recursively by the Levinson algorithm. This method is described, for example, in "Digital Signal Processing Theory 1 Basics / Control" (Corona).

FIGS. 7 and 8 show an embodiment in which the zero point is converted and smoothed in the time domain. In these figures, the deconvolution circuits 57 and 67 solve equation (15) for γ ′ (n) by calculating equation (17).

In FIG. 8, the output of the lag window 66 is subjected to FFT or DFT processing in place of the deconvolution circuit section 67, and the square reciprocal of the absolute value is obtained. To perform inverse FFT or inverse DFT processing. In this case, there is an effect that the calculation amount is further reduced as compared with the case of the inverse convolution.

 Next, the experimental results for real voice are shown below.

FIGS. 13 and 14 show the results of analysis of the “rain” uttered by one adult and one male. Audio input is performed by a μ-PCM codec with a sampling frequency of 8 kHz, and no preprocessing is performed. The analysis conditions include a frame of 30 ms (240 samples), an analysis length of 32 ms (256 samples), a Hamming window of 256 time windows, a pole order, and a zero order 8. FIGS. 13 (a) and 14 (a) show the spectrum of the all-pole model of the 16th order, the spectrum of the pole-zero model when spectrum smoothing is not performed, and the pole when the pitch period adaptation is smoothed. This is a comparison of the spectra of the zero model. As can be seen from these figures, when smoothing is not performed, false zeros or emphasized zeros appear in the spectrum of the pole-zero model,
The spectrum approximation is poor. However, when smoothing is performed, the zeros of the spectrum are well approximated. FIGS. 13 (b) and 14 (b) compare the methods of spectral smoothing. For male speech, there is no significant difference in the quality of the spectral approximation due to the difference in the smoothing method. This is considered to be because the interval of the spectral fine structure according to the pitch of the male voice used in the experiment was such that it could be smoothed by averaging three points. For female voices, the spectral approximation of the averaging method is poor. This is because in the case of female voice, the interval between fine structures became wide, and smoothing was not successfully performed by the averaging method. On the other hand, the peak-to-peak linear interpolation method has better spectral approximation than averaging because the range of smoothing changes according to the interval between fine structures. Table 1 shows the reciprocals of the segmental prediction gay Gseg and the segmental SFMseg when the smoothing method is changed. Gseg
And 1 / SFMseg are the predicted gain and 1 / SFM in each frame
It is the average in the B area. As can be seen from Table 1 and FIGS. 13 and 14, the pole-zero model using the pitch adaptation method as the spectrum smoothing method can always satisfactorily estimate the input spectrum regardless of the gender of the speaker.

As described above, according to the conventional method, due to the fine structure of the spectrum based on the periodicity of the voiced sound source, false zeros and emphasized zeros are generated in the spectrum of the residual signal of the all-pole model. There was a problem of incorrect parameter estimation. As a method for solving this problem, a method of smoothing the power spectrum of the residual signal is studied, and the power spectrum of the residual signal is smoothed in a frequency domain by a filter that adaptively changes a time constant according to a pitch period, afterwards,
A method to inverse spectrum and extract zero parameters was proposed. According to this method, parameters can always be extracted without error regardless of the fine structure of the spectrum. In addition, we clarified a method of performing spectrum smoothing and inverse spectrum conversion processing performed in the frequency domain in the time domain.

The proposed analysis method was applied to real speech, and it was shown that the zero of the spectrum could be approximated well.

Next, a principle and an embodiment of removing a spectrum and a fine structure by another attempt different from those shown in the above embodiment will be described with reference to FIGS. 15 to 20. FIG.

Modeling the sound generation mechanism as generally known results in a configuration as shown in FIG. That is, the pitch pulse e (t) (pitch T _p ) as the glottal signal
Is passed through a filter H (ω) as a vocal tract and an audio signal S (t) is output. In this case, the audio signal driven by the signal e ₀ (t) for one cycle of the pitch pulse e (t) is S ₀ (t).

The signal for one cycle of the pitch pulse is converted to E ₀ (t) as described above.
And its Fourier transform is E ₀ as shown in FIG. 17 (a).
(Ω), the Fourier transform E (ω) of e (t) is
As shown in FIG. 17 (b), E ₀ (ω) is set to 2π / T in the frequency direction.
_It is discretized for each _p . Therefore, the audio signal S
The spectrum S (ω) of (t) is also discretized (sampled) in the frequency direction as shown in FIG. 18 (c), and this is the fine structure of the spectrum generally referred to.

As described above, S ₀ (t) and S ₀ (t) are converted to an all-pole inverse filter (1 + A) as shown in FIG. 16 for the audio signal when driven by the signal e ₀ (t) for one pitch period. (omega)) d ₀ the residual signal when analyzed by all-pole model with (putting and t), the spectrum D ₀ (omega of d ₀ (t)), as shown in FIG. 19, S ( The pole is removed from ω), leaving zero.
On the other hand, the actual residual signal d (t) (FIG. 16) is a residual signal obtained by analyzing the audio signal S (t) driven and output by the periodic pitch pulse e (t). Therefore, the spectrum is obtained by sampling D ₀ (ω) in the frequency direction as shown in FIG. 19 (b).

Here, if the Fourier transform of D ₀ (ω) and D (ω) is FD ₀ (t), FD (t) as shown in FIGS. 20 (a) and 20 (b), FD (t) becomes FD (t). ₀ (t) becomes as that in the period signal having a period T _p. Therefore, if an ideal low-pass filter having a cutoff frequency T _p / 2 is applied to D (ω), D ₀ (ω) can be restored,
The fine structure due to the pitch can be removed from D (ω). Therefore, a low-pass filter having a cutoff frequency T _p / 2 may be used as the smoothing processing unit τ in FIG.

Also, The same applies to the case of the power spectrum, the power spectrum of the residual signal d (t) | by multiplying the cutoff frequency T _p / 2 of the low-pass filter ² | | H ^(ω) from ² | D ^(ω) Fine structure can be removed. Applying an ideal low-pass filter of | D (ω) | ² to T _p / 2 is equivalent to applying a square window of T _p / 2 to the autocorrelation coefficient of the residual signal in the time domain. Therefore, T _p / 2 as the lag window 66 in FIG.
May be used.

After removing the fine structure from the power spectrum of the residual signal by the above-described method, the zero parameter is obtained by applying the autocorrelation method to the autocorrelation coefficient obtained by the power spectrum reciprocal transform and inverse Fourier transform. As a result, the zero parameter can always be obtained with high accuracy without being affected by the pitch.

Note that the deconvolution process when the autocorrelation coefficient is multiplied by a square window of T _p / 2 is to solve the following matrix equation by Levinson's algorithm.

[Effects of the Invention] According to the present invention, it is possible to prevent a erroneous estimation of a zero point caused by a fine structure of a spectrum due to a pitch, and to obtain a pole-zero model that approximates a spectrum of an audio signal well.

[Brief description of the drawings]

FIG. 1 is a block diagram of a pole-zero model analysis method according to an embodiment of the present invention, FIG. 2 is a block diagram showing a specific example of a smoothing circuit which is a main part of FIG. 1, and FIG. FIG. 4 is a diagram showing an example of the spectrum of the obtained residual signal, FIG. 4 is a diagram showing an example of the frequency characteristic of the pole-zero model obtained by one embodiment of the present invention, and FIGS. FIG. 9 is a block diagram of a conventional pole-zero model analysis method according to another embodiment, FIG. 9 is a block diagram of a conventional pole-zero model analysis method, FIG. 10 is a diagram showing an example of a spectrum of a voice signal, and FIG. The figure showing an example of the spectrum of the residual signal that is the shape inverse filter output,
Fig. 12 shows an example of the frequency characteristics of the pole-zero model obtained by the conventional method, Figs. 13 and 14 show the analysis results for real speech, and Fig. 15 shows a model of voiced generation. Figure, No.
FIG. 16 shows a pole-zero analysis model, FIG. 17 shows a Fourier transform of a glottal signal (pitch pulse), and FIG.
_{(Ω), S 0 (ω} ), shows the characteristics of S (ω), Fig. 19 is D ₀
FIG. 20 shows spectra of (ω) and D (ω), and FIG.
FIG. 4 is a diagram showing a Fourier transform of ₀ (ω) and D (ω). 1,20,20,41,52,62,100 …… Input terminal, 2,31,42,53,62 …… Pole parameter estimation circuit, 3,32,43,54,63,102 …… All pole inverse filter, 4 , 33, 44… FFT circuit, 5, 34, 45… square circuit, 6, 35, 46, 56, 65… pitch analysis circuit, 7, 37… smoothing circuit, 8, 36, 47, 104… Reciprocal circuit, 9,38,48 …… Inverse FFT circuit, 10,39,50,59,68 …… Zero parameter estimation circuit, 11,40,51,60,69,107 …… All-pole inverse filter, 21,23 …… Multiplier, 22… Adder, 24 …… Unit delay circuit, 49,58,66 …… Lag window, 55,64 …… Autocorrelation coefficient calculator, 57,67 …… Deconvolution circuit, 101,106: Autocorrelation method execution circuit, 103: Power spectrum calculator.

Claims (3)

(57) [Claims] The present invention estimates the parameters of a pole model of an input speech signal, calculates the power spectrum of a residual signal obtained through an all-pole filter having the parameters of the pole model, and calculates the power spectrum of the power spectrum. In a voice signal analysis method for performing a pole-zero model analysis for obtaining a parameter of a zero model from an autocorrelation coefficient obtained by performing an inverse Fourier transform on a reciprocal, according to a pitch period value of an input voice signal or a residual signal. The power spectrum of the residual signal or its reciprocal is smoothed in the time domain. 2. An autocorrelation processing means for performing autocorrelation processing for estimating a polar model parameter of an input speech signal, and an all-pole inverse filter configured based on the polar model parameter. First filter means for obtaining a residual signal from the input audio signal, and arithmetic means for estimating the parameters of the zero model by performing autocorrelation processing on the result of inverse Fourier transform of the reciprocal of the power spectrum of the residual signal In the audio signal analysis method, the arithmetic unit includes a unit that extracts a pitch period of the input audio signal, and calculates a characteristic of a signal including a power spectrum of a residual signal based on a signal related to the pitch period. An audio signal analysis method, comprising means for smoothing a region. 3. A second method for obtaining a residual signal from an output signal of the first filter means via an all-zero type inverse filter formed based on the parameters of the zero model estimated by the arithmetic means.
3. The audio signal analysis method according to claim 2, further comprising a filter means for obtaining a spectrum envelope characteristic of the audio signal.