JP2844672B2 - Vocal vocal tract type speech analyzer - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vocal tract vocal tract type voice analyzer, and more particularly, to a vocal tract vocal tract type voice analyzer that simultaneously analyzes a vocal vocal source wave and characteristics of a vocal tract.

(Prior Art) There are three typical methods for analyzing a speech signal for speech synthesis and speech recognition. The first conventional method is based on a linear prediction method (LPC method) or ARMA analysis, and a speech signal is generated as a response to an impulse train for voiced sounds and a white noise for unvoiced sounds to a linear filter. This is a method of obtaining the coefficients of the linear filter based on a model that is performed. Various variants of this method depending on the construction of the linear filter include α (alpha) parameters (sometimes called AR functions), PARCOR (Percoll: partial autocorrelation) coefficients, ARMA (AIR or ALMA) A method of obtaining a coefficient called a coefficient or the like is known. These are Jay. Dee. Markel and A. H. Gray (JDMarkel and AHGray), "Linear Prediction of Speech (Translated by Suzuki)" (LINEAR Prediction
of Speech).

As a second conventional method, p. Hedeerin (P.Hedeli
n) by the International Conference
On Acoustic Speech Signal Processing (International Conference on Acoustics)
Speech Signal Processing International Conference on Audio and Speech Signaling '86, pp. 465 to 468 of the proceedings, High Quality Glottal LPC-Vocoding (High Quality Glottal LPC-Vocoding) There was something announced.

As shown in FIG. 3, this is based on a voice generation model in which a vocal cord sound source generated from the vocal cord sound source 101 instead of an impulse train is input to the linear filter 102 in the case of a voiced sound. Then, the voice generated by this model and the actual natural voice (input from the signal line 103)
Are optimized to minimize the square error (calculated at 104) of the sound source waveform, and the parameters of the sound source waveform and the AR coefficient are obtained.

As a third conventional example, a further improvement was made at the same conference in 1988 by Hedelin, who changed from 339 pages to 342 pages of a collection of papers in a phase.
Compensation in All Pole Speech
Analysis (Phase Compensation in All-Pole Speech
Analysis: phase correction in all-pole speech analysis).

In this method, a finite number of codes corresponding to various waveforms are prepared in advance, and for a vocal cord wave, an optimum one is selected and analyzed. That is, a limited number of sound source waveforms are prepared and numbered. In the analysis, first, a waveform of a certain code (number) is selected as a sound source, a voice is synthesized, and an error from an actual voice is obtained. This is repeated for waveforms of different codes, and the waveform of the code with the smallest error is used as the estimated vocal cord sound source waveform.

(Problem to be Solved by the Invention) According to the first conventional method described above, voice analysis can be performed with a small amount of calculation. However, when applied to voice synthesis or the like, synthesized voices with very good sound quality cannot be obtained for voiced sounds.
This is considered to be problematic because the impulse train is assumed as the sound source as described above. Physiological studies have revealed that vocal cord vibrations are very complex and cannot be approximated by simple impulses. This means that the characteristics represented by the analyzed α parameter, ARMA coefficient, and the like include characteristics of both the vocal tract sound source and the vocal tract, not the characteristics of the vocal tract.

The fact that the impulse train is not a good approximation of the actual sound source and the analysis results do not represent the characteristics of the vocal tract is not necessarily a fatal defect in coding for voice communication. It is possible to some extent to improve the sound quality by increasing the number of pulses. However, in speech synthesis of an arbitrary word that synthesizes speech from a character string in accordance with rules, it is necessary to independently control sound source parameters such as pitch and vocal tract parameters. A composite model is desired.
Also, in the coding, the efficiency and the sound are improved by analyzing based on a more accurate model.

The one that brings this point closer to actual speech generation is the second
And the third conventional method. That is, as a voiced sound source,
It uses a model of the waveform with reference to the result of observing the actual vocal cord vibration using a scope or the like. According to these conventional methods, naturalness of voiced sound and coding efficiency can be improved.

However, these second and third conventional methods aim at optimizing the coefficients while repeatedly evaluating the error, and therefore require enormous calculations for analysis, which is difficult to realize from the viewpoint of price and apparatus scale. There was a problem. This is because the analysis method such as the conventional linear prediction method cannot be applied even to the coefficients of the linear filter part having the same configuration as the first conventional method. In addition, it is necessary to match the position of the vocal cord wave on the time axis. Normally, the time length of the analysis frame is set to at least about three times the pitch, so that it is necessary to determine the positions of a plurality of vocal cord pulses between them, which further complicates the analysis.

An object of the present invention is to provide the above-described second and third conventional methods,
It is an object of the present invention to provide a vocal vocal tract-type speech analyzer that can be realized with a smaller amount of calculation while performing analysis based on a waveform model of a vocal fold sound source.

(Means for Solving the Problems) In order to solve the above problems, a vocal vocal tract type voice analysis device according to a first aspect of the present invention comprises: an input buffer for temporarily storing an input voice signal; FIR filter having the inverse characteristic of the above, coefficient generating means for generating the coefficient of the FIR filter having the inverse characteristic of the frequency characteristic of various vocal fold source waves, and spectrum analysis means for analyzing the spectrum of the output signal of the FIR filter And error comparison means for comparing the analysis errors sequentially obtained by the spectrum analysis means with each other, and when the input audio signal is input to the FIR filter having the coefficients of the FIR filter for the various vocal source sound wave candidates, The candidate of the vocal cord source wave when the analysis error obtained by the error comparing means is minimized is regarded as the analysis result of the vocal cord source wave, and the spectrum at that time is obtained. Control means for controlling to output the result of the torque analysis as the analysis result of the vocal tract characteristics.

Further, in the vocal vocal tract type voice analysis device according to the second invention of the present invention, an input buffer for temporarily storing an input voice signal, an IIR filter having the buffer output as an input, and a frequency characteristic of various vocal vocal source waves Generates the first coefficient of an IIR filter having the characteristic of the minimum phase component of the inverse characteristic of the above and the second coefficient of the IIR filter having the characteristic obtained by inverting the maximum phase component of the inverse characteristic of the frequency characteristic of the vocal cord source wave. Coefficient generating means for temporarily storing the output of the IIR filter, a time axis inverting buffer for inverting the time axis and outputting to the IIR filter, and a spectrum analyzing means for analyzing the spectrum of the output signal of the IIR filter, Error comparison means for comparing analysis errors sequentially obtained by the spectrum analysis means with each other; and for the various voice sound source candidates, the input sound signal is converted to the first coefficient. The time axis of the signal filtered by the IIR filter is inverted by the time axis inversion buffer, and the signal obtained by filtering the signal whose time axis is inverted again by the IIR filter having the second coefficient is converted to a time signal. The time axis is inverted again by the axis inversion buffer, and the analysis error when the signal is analyzed by the spectrum analysis means is compared with each other. When the analysis error is minimized, the candidate of the vocal fold source wave is determined as the vocal fold source wave. Control means for controlling the analysis result and outputting the result of the spectrum analysis at that time as the analysis result of the vocal tract characteristics.

(Operation) In the first embodiment of the present invention, first, a filter having the inverse characteristic of the frequency characteristic of the vocal cord source wave is realized by the FIR filter. Here, the inverse characteristic means that the amplitude characteristic has a reciprocal relationship, and the phase characteristic is obtained by reversing the positive and negative signs. This inverse characteristic
The coefficients of the FIR filter are obtained by calculating the Fourier transform of the original vocal cord wave, obtaining the inverse of the amplitude and the sign of the phase, and performing the inverse Fourier transform. As in the third conventional example, in the method of selecting the optimum one from the coded vocal folds, the coefficient of the FIR filter having the inverse characteristic may be obtained in advance for the vocal folds of each code. However, the amount of calculation at the time of analysis does not increase.

By the way, if the characteristic of the filter having the inverse characteristic is truly the reverse of the characteristic of the vocal fold wave, the characteristic of the signal passed through the filter having the inverse characteristic of the vocal fold wave is canceled by the characteristic of the vocal fold wave. Thus, the characteristic of only the vocal tract is represented. Therefore, the characteristics of the vocal tract can be estimated by analysis using the linear prediction method of the first conventional example. However, since the characteristics of the true vocal fold waves cannot be known in advance, it is not always possible to correctly analyze them with a single calculation. Therefore, the analysis is repeated assuming various vocal fold waves, and the one with the smallest error is determined as the optimal analysis result.

Further, in the second invention, an inverse filter having an inverse characteristic of the frequency characteristic of the vocal cord source wave is realized by the IIR filter. However, since the phase characteristic of the vocal cord wave cannot be represented by the minimum phase, if an inverse filter is to be realized by an IIR filter, it is necessary to set the pole outside the unit circle, which results in an unstable circuit. That is, simply the first
It is not possible to replace the FIR filter of the invention of the invention with an IIR filter. Therefore, the characteristics of the poles outside the unit circle are realized by performing filtering by inverting the time axis.

For this purpose, a time axis inversion buffer for temporarily storing the output of the IIR filter and inverting and outputting the time axis is provided. Then, filtering is performed by the IIR filter using the minimum phase component, that is, a coefficient corresponding to the pole zero inside the unit circle, and the output is time-axis inverted by the time-axis inversion buffer. The time-axis-inverted signal is filtered again by a characteristic corresponding to the characteristic obtained by inverting the maximum phase component, that is, a coefficient corresponding to the pole zero mapped inside the unit circle by replacing the pole zero outside the unit circle with the reciprocal, Further, the time axis is reversed again to return to the normal time axis. Although the process of inverting and filtering the time axis cannot be realized with an infinitely long signal, the present invention can be realized by processing the input audio signal by dividing it into analysis frames having a finite inter-axis length. Subsequent steps are the same as in the first invention.

In these inventions, it is not necessary to estimate the position of the sound source wave, and the vocal tract characteristics can be obtained as values of the α parameter or the like by using a high-speed analysis method such as the first conventional example.
In this way, the values of the parameters representing the characteristics of the vocal cords and the vocal tract are obtained. The repetitive search for the evaluation of the error for various sound source waves is the same as in the second and third conventional examples, but the amount of calculation for each repetition is small.

It is to be noted that the characteristics such as the α parameter obtained as a result do not include the characteristics of the vocal cords, as in the case of the first conventional example, and are considered to closely approximate the characteristics of the vocal tract. Therefore, by realizing the inverse characteristic of this characteristic and filtering the original audio signal (this is a method called inverse filtering), a vocal cord source wave as a function of time is obtained. This means that not only the shape of the vocal cord waves but also the position information on the time axis can be obtained. The second and third
Unlike the prior art example in which the position must first be estimated at the same time, if the analysis is completed and inverse filtering is performed, the estimation of the sound source waveform including the position information can be easily performed.

(Example) Next, a first example of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing one embodiment of the first invention of the present invention.

The input buffer 2 temporarily stores the audio signal input from the signal line 9 and sends the audio signal for one analysis frame to the FIR filter 8 according to the control signal from the control circuit 1.

The vocal cord candidate setting circuit 6 generates a vocal cord candidate code and sends it to the FIR coefficient memory 7 and the analysis result buffer 4. As a code generation method, if a code is searched for all codes in a round robin search, the code may be generated in order.
Alternatively, 0 may be set.

The FIR coefficient memory 7 previously stores the coefficient values of the FIR filter having the inverse characteristics of the frequency characteristics of various vocal fold source waves, and when the vocal fold wave candidate setting circuit 6 sends the code of the vocal fold wave candidate. , The coefficient value of the corresponding FIR filter
Send to filter 8.

The FIR filter 8 filters the audio signal sent from the input buffer 2 using the coefficient value sent from the FIR coefficient memory 7 and sends it to the linear prediction analysis circuit 3.

The linear prediction analysis circuit 3 performs a linear prediction analysis of the signal sent from the FIR filter 8 and sends a linear prediction coefficient (α parameter or PARCOR coefficient) to the analysis result buffer 4.
The analysis error is sent to the error comparison circuit 5.

Analysis result buffer 4 is a double buffer,
The first buffer temporarily stores the vocal fold wave candidate codes sent from the vocal fold wave candidate setting circuit 6 and the linear prediction coefficients sent from the linear prediction analysis circuit 3, and the second buffer stores the vocal fold wave candidate codes. The code and the candidate of the optimal value of the linear prediction coefficient are stored. When a control signal indicating an instruction to update the optimum value is sent from the control circuit 1, the contents of the first buffer are copied to the second buffer. Thus, when the analysis of all the vocal fold wave candidates is completed, the data remaining in the second buffer becomes the optimal vocal fold wave signal and the data of the linear prediction coefficient.

The error comparison circuit 5 has a memory and a comparator for storing a candidate for the minimum value of the analysis error, compares the analysis error in the memory with the analysis error sent from the linear prediction analysis circuit 3, and determines that the latter is smaller. , The information in the memory is rewritten and the control circuit 1 is notified. Note that the memory is reset at the beginning of each analysis frame.

The control circuit 1 causes the vocal fold wave candidate setting circuit 6 to generate a vocal fold wave candidate code, and sends it to the FIR coefficient memory 7 and the analysis result buffer 4. Next, the FIR filter 8 is caused to set the coefficient value sent from the FIR coefficient memory 7 as a filter coefficient.
Then, the signal for one analysis frame is input to the input buffer 2 by FIR.
Let the filter 8 send. Then, the filtered signal is analyzed by the linear prediction analysis circuit 3, and the linear prediction coefficient and the analysis error are analyzed by the analysis result buffer 4 and the error comparison circuit 5, respectively.
To send. When the error comparison circuit 5 notifies that the new analysis error is smaller, it instructs the analysis result buffer 4 to update the optimum value. This is repeated for various sound source candidates.

When the analysis of all the vocal fold wave candidates is completed, the analysis result buffer 4 outputs the data remaining in the second buffer from the signal line 10.

 The above operation is repeated for each analysis frame.

Next, the second invention will be described with reference to FIG.

In the figure, an input buffer 2, a linear prediction analysis circuit 3,
The analysis result buffer 4, error comparison circuit 5, and vocal cord candidate setting circuit 6 operate in the same manner as in the first embodiment.

The IIR coefficient memory 11 stores the first coefficient of the IIR filter having the characteristic of the minimum phase component of the inverse characteristic of the frequency characteristic of the vocal cord source wave and the maximum phase component of the inverse characteristic of the frequency characteristic of the vocal cord source wave in the opposite phase. IIR filter with generalized characteristics
Are stored. The value of this coefficient is transmitted to the IIR filter in accordance with the sign of the sound source wave candidate sent from the vocal cord candidate setting circuit 6 and the control signal indicating whether the signal is the first coefficient or the second coefficient sent from the control circuit 1. Can be

These coefficients can be created in advance by the following processing. First, a complex cepstrum of a sound source wave is obtained. The component of the time origin is replaced with a value obtained by subtracting 1 from the value, and the signs of the components other than the time origin are inverted. The complex cepstrum obtained in this way corresponds to the inverse characteristic of the characteristic of the sound source wave. The positive time portion corresponds to the minimum phase component. For example, a filter coefficient for realizing this can be easily obtained by obtaining a power spectrum from the Fourier transform, obtaining an autocorrelation function by the inverse Fourier transform, and using a linear prediction method. If this coefficient is used as the coefficient of the IIR filter, a filter having the minimum phase component of the inverse characteristic of the vocal fold source wave can be realized. That is, this is the first coefficient. On the other hand, if the same processing is performed by inverting the time axis of the negative time portion of the complex cepstrum having the inverse characteristic of the vocal fold source wave, the second coefficient is obtained.

The IIR filter 12 filters a signal sent from the input buffer 2 or the time axis inversion buffer 13 and outputs the filtered signal to the time axis inversion buffer 13. Switching of an input signal to the filter is performed by a switch according to a control signal from the control circuit 1.
Done by 14.

The time axis inversion buffer 13 temporarily stores the signal sent from the IIR filter 12, inverts the time, and outputs the inverted signal. This signal is sent to the IIR filter 12 or the linear prediction analysis circuit 3 by switching the switch 15 with a control signal from the control circuit 1.

The control circuit 1 first causes the vocal fold wave candidate setting circuit 6 to generate a vocal fold wave candidate code, and sends it to the IIR coefficient memory 11 and the analysis result buffer 4. At the same time, the IIR coefficient memory 11 has the first
Is output. Next, the IIR filter 12
Causes the coefficient value sent from the IIR coefficient memory 11 to be set as a filter coefficient. Subsequently, the switch 14 is switched to the input buffer 2 side, and the input buffer 2 sends a signal for one analysis frame to the IIR filter 12. The filtered signal is sent to the time axis inversion buffer 13. Next, it instructs the IIR coefficient memory 11 to output the second coefficient with respect to the code of the same excitation wave. The coefficient is converted to an IIR filter 12
, And switches the switches 14 and 15 so that the signal flows from the time axis inversion buffer 13 to the IIR filter 12, and the filtered signal is sent to the time axis inversion buffer 13 again. Next, the switch 15 is switched to send the output signal of the time axis inversion buffer 13 to the linear prediction analysis circuit 3 for analysis.

Thereafter, each unit is controlled in the same manner as in the embodiment shown in FIG. 1, and the analysis result buffer 4 outputs the optimum analysis result data remaining in the second buffer from the signal line 10.

In this embodiment, as compared with the embodiment shown in FIG. 1, an extra time axis inversion buffer is required, but the amount of calculation can be reduced because the IIR filter can be used.

(Effects of the Invention) As described above, the present invention is an analysis based on a waveform model of an audio sound source, but can be realized with a smaller amount of calculation. This has the effect that the included sound source waveform can be easily estimated.

[Brief description of the drawings]

FIG. 1 is a block diagram showing one embodiment of the first invention in the present invention, FIG. 2 is a block diagram showing one embodiment of the second invention in the present invention, and FIG. 3 is for explaining a conventional method. It is a block diagram of. 1 ... control circuit, 2 ... input buffer, 3 ... linear prediction analysis circuit, 4 ... analysis result buffer, 5 ... error comparison circuit, 6 ... vocal band candidate setting circuit, 7 ... FIR coefficient memory, 8 ... FIR filter, 11 ... IIR coefficient memory, 12 ...
IIR filter, 13: time axis inversion buffer, 14, 15, ... switch.

Claims (2)

(57) [Claims] A vocal vocal tract type voice analyzer for analyzing the characteristics of a vocal tract source wave and a vocal tract of a voice, comprising: an input buffer for temporarily storing an input voice signal; FIR filter having characteristics, coefficient generating means for generating coefficients of the FIR filter having the inverse characteristic of the frequency characteristics of various vocal source sound waves, spectrum analysis means for analyzing the spectrum of the output signal of the FIR filter, Error comparing means for comparing the analysis errors sequentially obtained by the spectrum analyzing means with each other, and comparing the error when the input voice signal is input to the FIR filter having the coefficients of the FIR filter for the various vocal source sound wave candidates. The candidate of the vocal cord source wave when the analysis error obtained by the means is minimized is taken as the analysis result of the vocal cord source wave, and the result of the spectrum analysis at that time is taken as the vocal tract. Sexual analysis as constituted by a control means for controlling to output the possible vocal cord vocal tract type speech analysis apparatus according to claim. 2. A vocal cord analyzing device of the type for analyzing the sound source wave of a voice and the characteristics of a vocal tract. An input buffer for temporarily storing an input voice signal, and an IIR having the buffer output as an input.
The first coefficient of the IIR filter, which has the characteristic of the minimum phase component of the inverse characteristic of the frequency characteristic possessed by the filter and the various vocal fold source waves, and the maximum phase component of the inverse characteristic of the frequency characteristic possessed by the vocal fold source wave are inverted. Coefficient generating means for generating a second coefficient of an IIR filter having characteristics; a time axis inversion buffer for temporarily storing an output of the IIR filter, inverting a time axis and outputting the inverted time axis to the IIR filter; Spectrum analysis means for analyzing the spectrum of the output signal; error comparison means for comparing the analysis errors sequentially obtained by the spectrum analysis means with each other; The IIR having a coefficient of 1
Invert the time axis of the signal filtered by the filter in the time axis inversion buffer, and then invert the time axis inverted signal of the signal whose time axis is inverted again by the IIR filter having the second coefficient. The time axis is inverted again in the buffer, and the analysis errors when the signals are analyzed by the spectrum analysis means are compared with each other. When the analysis error is minimized, the candidate of the vocal vocal source wave is analyzed for the vocal source wave. A vocal tract vocal tract type voice analysis device, comprising: a control unit for controlling the output of the result of the spectrum analysis at that time as a result of analysis of the vocal tract characteristics.