DE69024899T2 - Method and device for analysis by synthesizing speech - Google Patents

Description

The present invention relates to a method and a device for speech analysis synthesis, in which a linear filter representing the spectral envelope characteristic of a speech is excited to synthesize a speech signal by means of an excitation signal.

A linear predictive vocoder and a multi-pulse predictive coding have been proposed for use in speech analysis synthesis systems of this type. The linear predictive vocoder is now widely used for speech coding in a low bit rate range below 4.8 kb/s, and this system includes a PARCOR system and a line spectrum pair (LSP) system. These systems are described in detail, for example, in Saito and Nakata, "Fundamentals of Speech Signal Processing," ACADEMIC PRESS, INC., 1985. The linear predictive vocoder is constructed of an all-pole filter representing the spectral envelope characteristic of a speech and an excitation signal generator part for generating a signal for exciting the all-pole filter. The excitation signal is a pitch-frequency pulse train for a voiced sound and white noise for an unvoiced sound. Excitation parameters are the distinction between voiced and unvoiced sounds, the pitch frequency and the strength of the excitation signal. These parameters are extracted as averaged features of the speech signal in an analysis window of about 30 msec. In the linear predictive vocoder, since parameters extracted for each analysis window are temporally interpolated for speech synthesis as mentioned above, features of its waveform cannot be reproduced with sufficient accuracy when the pitch frequency, strength and spectral characteristics of the speech undergo rapid changes. Furthermore, since the excitation signal composed of the pitch frequency pulse train and the white noise is insufficient to reproduce features of various speech waveforms, it is difficult to produce a highly natural-sounding synthesized speech. To improve the quality of the synthesized speech in the linear predictive vocoder, the prior art has considered using an excitation that allows more accurate reproduction of the features of the speech waveforms.

On the other hand, predictive multi-pulse coding is a method that uses excitation of higher manufacturability than the conventional vocoder. In this method, the excitation signal is expressed using a plurality of pulses, and two all-pole filters representing the proximity correlation and pitch correlation characteristics of speech are excited by the excitation signal for speech synthesis. The timing and strength of the pulses are selected so that an error between the input original speech waveform and the synthesized speech waveform becomes minimal. This is described in detail in BS Atal, "A New Model of LPC Excitation for Producing Natural-Sounding Speech at Low Bit Rates," IEEE Int. Conf. on ASSP, pages 614-617, 1982. With multipulse predictive coding, the speech quality can be improved by increasing the number of pulses used, but when the bit rate is low, the number of pulses is limited, and consequently the reproducibility of the speech waveform is deteriorated and sufficient speech quality cannot be obtained. It is considered in the prior art that an information amount of about 8 kb/s is required to produce high speech quality.

In the predictive multipulse coding, the excitation is set so that the input speech waveform itself is reproduced. On the other hand, a method has also been proposed in which a phase-adjusted speech signal resulting from equalizing a phase component of the speech waveform with a certain phase is subjected to the predictive multipulse coding, as set forth in U.S. Patent No. 4,850,022 assigned to the inventor of the present application. This method improves the speech quality at low bit rates because the number of pulses for reproducing the excitation signal can be reduced by removing from the speech waveform the phase component of a speech that is not significant to human hearing. However, in this method, when the bit rate drops to 4.8 kb/s or so, the number of pulses for reproducing features of the speech waveform with high accuracy becomes insufficient, and high-quality speech cannot be produced either.

SUMMARY OF THE INVENTION

It is therefore an object of the present invention to provide a method and an apparatus for speech analysis synthesis which allow the generation of high quality speech at bit rates in the range of 2.4 to 4.8 kb/s, i.e. in the border area between the amounts of information required for the linear predictive vocoder and the speech waveform coding.

This object is achieved with a method according to claim 1, a speech analysis device according to claim or a speech synthesis device according to claim 9.

Preferred embodiments of the invention are the subject of the dependent claims.

According to the present invention, a quasi-periodic pulse train is generated with limited variation of its pitch period. By using the quasi-periodic pulse train as an excitation signal for a voiced sound in speech analysis, it is possible to reduce the amount of parameter information that the pulse train represents.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is a block diagram showing an embodiment of the speech analysis synthesis method according to the present invention;

Fig. 2 is a block diagram showing an example of a phase adjustment and analysis part 4 ;

Fig. 3 is a diagram for explaining a quasi-periodic pulse excitation signal;

Fig. 4 is a flowchart of a pulse position generating process;

Fig. 5A is a diagram for explaining the insertion of a pulse position in Fig. 4;

Fig. 5B is a diagram for explaining the shift of a pulse position in Fig. 4;

Fig. 5C is a diagram for explaining the distance of a pulse position in Fig. 4;

Fig. 6 is a block diagram showing an example of a pulse size calculation part 8;

Fig. 6A is a block diagram showing a frequency weighting filter processing part 39 shown in Fig. 6;

Fig. 7A is a diagram showing an example of the waveform of a phase-aligned prediction residual;

Fig. 7B is a graph showing an impulse response of a zero filter;

Fig. 8 is a block diagram showing an example of a zero filter coefficient calculation part 11;

Fig. 9 is a block diagram showing another example of the pulse magnitude calculation part 8 ; and

Fig. 10 is a graph showing the comparison results of synthesized speech quality between the present invention and the prior art.

DESCRIPTION OF THE PREFERRED EMBODIMENT

Fig. 1 shows in block form the structure of the speech analysis synthesis system of the present invention. A sampled digital speech signal s(t) is input via an input terminal 1. In a linear predictive analyzer section 2, the samples of N speech signals are stored at once in a data buffer for each analysis window and these Samples are then subjected to linear predictive analysis by a known linear predictive coding method to calculate a set of prediction coefficients ai (where i = 1, 2, ..., p). In the linear predictive analyzer part 2, a prediction residual signal e(t) of the input speech signal s(t) is obtained by means of an inverse filter (not shown) which uses the set of prediction coefficients as its filter coefficients. On the basis of the decision of the level for a maximum value of an autocorrelation function of the prediction residual signal, it is determined whether the speech is voiced (V) or unvoiced (U), and a decision signal UV is output accordingly. This processing is described in detail in the aforementioned reference by Saito et al. The prediction coefficients ai obtained in the linear predictive analyzer part 2 are supplied to a phase matching analyzer part 4 and quantized by a quantizer 3 at the same time.

In the phase equalization analyzer part 4, coefficients of a phase equalization filter which is to make the phase characteristic of the speech a zero phase and reference times of the phase equalization are calculated. Fig. 2 shows the detailed structure of the phase equalization analyzer part 4. The speech signal s(t) is applied to an inverse filter 31 to obtain the prediction residual e(t). The prediction residual e(t) is supplied to a maximum magnitude position detector part 32 and a phase equalization filter 37. A switch control part 33C monitors the decision signal VU supplied from the linear predictive analyzer part 2 and normally connects a switch 33 to the output side of a magnitude comparison part 38. However, when the current window is of a voiced sound V and the immediately preceding frame is of an unvoiced sound U, the switch 33 is connected to the output side of the maximum magnitude position detector part 32. In this case, the maximum magnitude position detector part 32 detects and outputs a sampling point t'p at which the magnitude of the prediction residual e(t) is maximum.

Assume that smoothed phase matching filter coefficients ht'i(k) for the currently determined reference time t'i have been obtained at a coefficient smoothing part 35. The coefficients ht'i(k) are supplied from the filter coefficient holding part 36 to the phase matching filter 37. The prediction residual e(t), which is the output of the inverse filter 31, is phase-matched by the phase matching filter 37 and outputted therefrom as a phase-matched prediction residual ep(t). It is known that when the input speech signal s(t) is a voiced sound signal, the prediction residual e(t) of the speech signal has a waveform having pulses at the pitch intervals of the voiced sound. The phase equalization filter 37 produces an effect of emphasizing the magnitudes of the pulses of such pitch intervals.

The magnitude comparison part 38 compares levels of the phase-adjusted prediction residual ep(t) with a predetermined threshold value, determines, as a pulse position, each sampling time point where the sampling value exceeds the threshold value, and outputs the pulse position as the next reference time point t'i+1 under the condition that an allowable minimum value of the pulse intervals is Lmin, and the next reference time point t'i+1 is searched for among sampling points spaced more than the value Lmin from the time point t'i.

If the frame is an unvoiced phone frame, the phase-matched residual ep(t) during the unvoiced phone frame is composed essentially of random components (or white noise) that are significantly lower than the above-mentioned threshold, and the magnitude comparison part 38 does not generate the next reference time point t'i+1 as an output of the phase-matching analyzer part 4. Instead, the magnitude comparison part 38 determines a dummy reference time point t'i+1 at, for example, the last sampling point of the frame (but not limited to), for use in determining the smoothed filter coefficients at the smoothing part 35, as explained later.

In response to the next reference time t'i+1 thus obtained in the voiced sound frame, a filter coefficient calculation part 34 calculates (2M + 1) filter coefficients h*(k) of the phase adjustment filter 37 according to the following equation:

where k = -M, -(M-1), ..., 0, 1, ..., M.

On the other hand, when the frame is an unvoiced phone frame, the filter coefficient calculation part 34 calculates the filter coefficients h*(k) of the phase adjustment filter 37 using the following equation:

where k = -M, ..., M.

The characteristic of the phase matching filter 37 expressed by equation (2) represents such a characteristic that its input signal is passed through unchanged.

The filter coefficients h*(k) thus calculated for the next reference time ti+1 are smoothed by the coefficient smoothing part 35 as described later to obtain smoothed phase-adjustment filter coefficients ht'i+1(k), which are held by the coefficient holding part 36 and supplied as renewed coefficients ht'i(k) to the phase-adjustment filter 37. The phase-adjustment filter 37, whose coefficients have been renewed in this way, again performs phase-adjustment of the prediction residual e(t), and on the basis of its output, the next pulse position, that is, a new next reference point t'i+1, is determined by the magnitude comparing part 38. In this way, a next reference time t'i+1 is determined on the basis of the phase-adjusted residual ep(t) output from the phase-adjustment fitter 37 whose coefficients have been set to ht'i(k), and then, new smoothed filter coefficients hti+1(k) are calculated for the reference time t'i+1. By repeating these processes using the reference time t'i+1 and the smoothed filter coefficients ht'i+1(k) as new t'i and ht'i(k), reference time points in each frame and the smoothed filter coefficients ht'i(k) for these reference time points are determined in sequential order.

In the case where speech begins after a period of silence or a voiced sound begins after continued unvoiced sounds, the prediction residual e(t) containing pulses of the pitch frequency is supplied for the first time to the phase matching filter 37 in which the filter coefficients given essentially by equation (1) are set. In this case, the magnitudes of the pulses are not emphasized and hence the prediction residual e(t) is output unchanged from the filter 37. Thus, if the magnitudes of the pulses of the pitch frequency happen to be smaller than the threshold value, the pulses cannot be detected in the magnitude comparison part 38. That is, the speech is processed as if no pulses were included in the prediction residual and hence the filter coefficients h*(k) for the pulse positions are not obtained. This is unfavorable from the point of view of speech quality in speech analysis synthesis.

To solve this problem, in the embodiment in Fig. 2, when the input speech signal analysis window changes from an unvoiced phone frame to a voiced phone frame as mentioned above, the maximum magnitude detector part 32 detects the maximum magnitude position t'p of the prediction residual e(t) in the voiced phone frame and supplies it to the filter coefficient calculation part 34 via the switch 33 and simultaneously outputs it as a reference time point. The filter coefficient calculation part 34 calculates the filter coefficients h*(k) using the reference time point t'p instead of t'i+1 in equation (2).

Next, a description will be given of the smoothing process of the phase matching filter coefficients h*(k) by the coefficient smoothing part 35. The filter coefficients h*(k) determined for the next reference time t'i+1 and supplied to the smoothing part 35 are temporarily smoothed by a first-order filtering process expressed, for example, by the following recursion formula:

ht(k) = bht-1(k) + (1-b)h*(k) ...(3)

where: t'i < t &le; t'i+1.

The coefficient b is set to a value of about 0.97. In equation (3), ht-1(k) represents smoothed filter coefficients at any sampling point (t-1) in the time interval between the current reference time t'i and the next reference time t'i+1, while ht(k) represents the smoothed filter coefficients at the next sampling point. This smoothing is performed for each sampling point from a sampling point next to the current reference time t'i for which the smoothed filter coefficients have already been obtained to the next reference time t'i+1 for which the smoothed filter coefficients are to be obtained next. The filter coefficient holding part 36 holds those of the filter coefficients thus sequentially smoothed ht(k) obtained for the last sampling point which is the next reference time, ie ht'i+1(k), and supplies them as renewed filter coefficients ht'i(k) to the phase adjustment filter 37 for further determination of a subsequent next reference time.

The phase adjustment filter 37 is supplied with the prediction residual e(t) and calculates the phase-adjusted prediction residual ep(t) using the following equation:

The calculation of the equation (4) only needs to be carried out until the next pulse position is detected by the magnitude comparison part 38 after the reference time t'i at which the above-mentioned smoothed filter coefficients were obtained. In the magnitude comparison part 38, the magnitude level of the phase-adjusted prediction residual ep(t) is compared with a threshold value, and the sampling point where the former exceeds the latter is determined as the next reference time t'i+1 in the current frame. Incidentally, in the case where no magnitude exceeds the threshold value within a predetermined period after the last pulse position (reference time) t'i, processing is carried out by which the time at which the phase-adjusted prediction residual ep(t) assumes the maximum magnitude up to that point is determined as the next reference time t'i+1.

The procedure for obtaining the reference time t'i and the smoothed filter coefficients ht'i(k) at this point, as described above, can be briefly summarized in the following overview.

Step 1: First, the phase-aligned prediction residual ep(t) is calculated using equation (4) using the filter coefficients ht'i(k) set up until then in the phase-alignment filter 37, i.e. the smoothed filter coefficients obtained for the last pulse position in the previous frame, as well as the prediction residual ep(t) of the given frame. This calculation only needs to be carried out until the detection of the next pulse after the previous pulse position.

Step 2: The magnitude of the phase-adjusted prediction residual is compared with a threshold value in the magnitude comparison part 38, the sampling point at which the residual exceeds the threshold value is determined as a pulse position, and the first pulse position t'i+1 (i = 0, i.e., t₁) in the current frame is obtained as the next reference time point.

Step 3: The coefficients h*(k) of the phase matching filter at the reference time t₁ are calculated, where the time t₁ is substituted for t'i+1 in equation (1).

Step 4: The filter coefficients h*(k) for the first reference time t₁ is inserted into equation (3) and the smoothed filter coefficients ht(k) at each of sampling points after the previous pulse position (the last pulse position t₀ in the previous frame) are calculated from equation (3) until the time of the pulse position t₁. The smoothed filter coefficients at the reference time t₁ obtained as a result are represented by ht(k).

Step 5: The phase-adjusted prediction residual ep(t) is calculated by substituting the smoothed filter coefficients hti(k) for the reference time t1 into equation (4). This calculation is performed for a period from the reference time t1 until the detection of the next pulse position (reference time) t2.

Step 6: The second pulse position t2 of the phase-aligned prediction residual thus calculated is determined in the magnitude comparison part 38.

Step 7: The second pulse position t2 is substituted for the reference time t'i+1 in equation (1) and the phase-matched filter coefficients h*(k) for the pulse position t2 are calculated.

Step 8: The filter coefficients for the second pulse position t₂ are substituted into equation (4), and the smoothed filter coefficients at respective sampling points are calculated in order, starting from a sampling point closest to the first pulse position t₁ and ending at the second pulse position t₂. As a result, the smoothed filter coefficients ht₂(k) at the second pulse position t₂ are obtained.

Thereafter, for example, steps 5 to 8 are repeatedly performed in the same manner as mentioned above, whereby the smoothed filter coefficients h*(k) can be obtained at all pulse positions in the frame.

As shown in Fig. 1, the smoothed filter coefficients ht(k) obtained in the phase equalization analyzer part 4 are used to control the phase equalization filter 5. By inputting the speech signal s(t) to the phase equalization filter 5, the processing expressed by the following equation is carried out to obtain a phase-adjusted speech signal Sp(t).

Next, an excitation parameter analyzer part 30 will be described. In the analysis-synthesis method of the present invention, different excitation sources for voiced and unvoiced sounds are used, and a switch 17 is switched by the voiced or unvoiced sound decision signal VU. The excitation source for the voiced sound comprises a pulse train generator part 7 and an all-zero filter (hereinafter simply referred to as a zero filter) 10.

The pulse train generating part 7 generates such a quasi-periodic pulse train as shown in Fig. 3 in which the pulse position ti and the size mi of each pulse are specified. The temporal position (the pulse position) ti and the size mi of each pulse in the quasi-periodic pulse train are represented as parameters. The pulse position ti is generated by a pulse position generating part 6 on the basis of the reference time t'i, and the pulse size m is controlled by a pulse size calculating part 8.

In the pulse position generator part 6, the interval between the reference time points (representing the positions of pulses of the pitch frequency in the phase-aligned prediction residual) determined in the phase-alignment analyzer part 4 is controlled to be quasi-periodic in order to reduce fluctuations in the pulse position and to reduce the amount of information required to represent the pulse position. That is, the interval Ti = ti - ti-1 between pulses to be generated shown in Fig. 3 is limited so that a difference in the interval between successive pulses is equal to or smaller than a fixed allowable value J as expressed by the following equation:

ΔTi = Ti - Ti-1 ≤ J ...(6)

Next, an example of the pulse position generating procedure that the pulse position generating part 6 implements will be described with reference to Fig. 4.

Step S₁: When all the reference time points t'i (where i = 1, 2, ...) in the current frame are input from the phase matching analyzer part 4, the process proceeds to the next step S₂ if the previous frame is a voiced phone frame (the current frame is also a voiced phone frame).

Step S2: A difference ΔT1 = Ti - Ti-1 is calculated between two consecutive intervals Ti = t'i - ti-1 and Ti-1 = ti-1 - ti-2 of the first reference instant ti (where i = 1) and the two pulse positions ti-1 and ti-2 already determined by the processing of Fig. 4 (already determined by the processing for the last two reference instants ti-2 and ti-1 in the previous frame).

Step S₃: The absolute value of the difference ΔT₁ is compared with the predetermined value J. If the former is equal to or smaller than the latter, it is determined that the inputted reference time t'i. is within a predetermined range of variation, and the process proceeds to step S₄. If the former is larger than the latter, it is determined that the reference time t'i changes beyond the predetermined limit, and the process proceeds to step S₆.

Step S4: Since the reference time t'i. is within the predetermined variation range, this reference time is determined as the pulse position ti.

Step S₅: It is determined whether or not the processing for all the reference time points t'i in the frame is completed, and if not, the process goes back to step S₂ and starts the processing for the next reference time point ti+1. If the processing for all the reference time points is completed, then the process goes to step S₁₇.

Step S6: A difference ΔT2 = (t'i - ti-1)/2 - (ti-1 - ti-2) is calculated between half of the interval Ti between the pulse position ti-1 and the reference time t'i and the already determined interval Ti-1.

Step S7: The absolute value of the above-mentioned difference ΔT2 is compared with the value J, and if the former is equal to or smaller than the latter, the interval T is about twice as large as the decided interval Ti-1 as shown in Fig. 5A. In this case, the process proceeds to step S8.

Step S�8: A pulse position tc is set approximately halfway between the reference time t'i and the previous pulse position ti-1, and the reference time t'i is set to the pulse position ti+1, and the process then proceeds to step S�5; .

Step S9: If the condition in step S7 is not satisfied, a calculation is made of a difference ΔT3 between the interval from the next reference time ti+1 to the pulse position ti-1 and the decided interval from the pulse position ti-1 to ti-2.

Step S₁₀: The absolute value of the above-mentioned difference ΔT₃ is compared with the value J. If the former is equal to or smaller than the latter, the reference time t'i+1 is within an expected range of the pulse position ti closest to the decided pulse position ti-1 and the reference time t'i is outside the range and between t'i+1 and ti-1. The process proceeds to step s₁₁.

Step S₁₁: The outstanding reference time t'i shown in Fig. 5B is disregarded, and instead the reference time t'i+1 is set at the pulse position ti, and the process proceeds to step S₅.

Step S₁₂: If the condition in step S₁₀ is not satisfied, a calculation is made of a difference ΔT₄ between half of the interval between the reference time t'i+1 and the pulse position ti-1 and the above-mentioned decided interval Ti-1.

Step S₁₃: The absolute value of the difference ΔT₄ is compared with the value J. If the former is equal to or smaller than the latter, it means that the reference time t'i+1 is within an expected range of the pulse position t'i+1 closest to that of ti as shown in Fig. 5C, and that the reference time t'i is one of two reference times t'i shown in Fig. 5C and is outside an expected range of the pulse position ti. In this case, the process proceeds to step S₁₄.

Step S₁₄: The reference time t'i+1 is set as the pulse position ti+1, and at the same time the reference time t'i is shifted to the middle between t'i+1 and ti-1 and set as the pulse position ti, that is, ti = (t'i+1 + ti-1)/2. The process proceeds to step S₅.

Step S₁₅: If the condition in step S₁₄ is not satisfied, the reference time t'i is set as the pulse position ti without taking any step because of its unsuitability as a pitch position. The process proceeds to step S₅.

Step S₁₆: If in step S₁ the previous frame is an unvoiced phone frame, all reference time points t'i in the current frame are set to the pulse positions ti.

Step S17: The number of pulse positions is compared with a predetermined maximum allowable number of pulses Np, and if the former is equal to or smaller than the latter, the entire processing is terminated. The number Np is a fixed integer in the range of, for example, 5 to 6, and this is the number of pulses present in a 15 msec frame in the case where the upper limit of the pitch frequency of a speech is considered to range from about 350 to 400 Hz at most.

Step S₁₈: When the condition in step S₁₇ is not satisfied, the number of pulse positions is larger than the number Np, so that sizes of pulses for the respective pulse positions are calculated by the pulse size calculation part 8 in Fig. 1, as described later.

Step S₁₉: A pulse position selecting part 6A in Fig. 1 selects Np pulse positions in the order of magnitude and indicates the selected pulses to the pulse position generating part 6, thus terminating the process.

According to the processing described above with reference to Fig. 4, even if the pulse position of the phase-matched prediction residual determined as the reference time t'i undergoes a substantial change, the fluctuation of the pulse position ti generated by the pulse position generating part 6 is limited to within a certain range. Thus, the amount of information required to represent the pulse position can be reduced. Moreover, even in the case where the pulse magnitude at the pitch position in the phase-matched prediction residual should be smaller than a threshold value and cannot be detected by the magnitude comparing part 38 in Fig. 2, a pulse signal is detected by the Steps S�7 and S�8 in Fig. 4 are inserted so that the quality of the synthesized speech is not significantly deteriorated despite an error in pulse detection.

In the pulse size calculation part 8, the pulse size at each pulse position t1 generated by the pulse position generator part 6 is selected so that a frequency-weighted mean square error between a synthesized speech waveform Sp'(t) generated by exciting such an all-pole filter 18 with the pulse train generated by the pulse train generator part 7 and an input speech waveform Sp(t) phase-adjusted by a phase-adjustment filter 5 can be finally minimized. Fig. 6 shows the internal structure of the pulse size calculation part 8. The phase-adjusted input speech waveform Sp(t) is supplied to a frequency-weighting filter processing part 39. The frequency-weighting filter processing part 39 expands the bandwidth of the resonance frequency components of a speech spectrum, and its transfer characteristic is as follows:

Hw(z) = A(z)/A(z/γ) ... (7)

where:

A(z) = 1 + a₁z⊃min;¹ + ... + apz-p ... (8)

where ai are linear prediction coefficients and z-1 is a sampling delay. γ is a parameter that controls the suppression level and is in the range of 0 < γ ≤ 1. The suppression level increases with decreasing value of γ. Usually γ is in the range of 0.7 to 0.9.

The frequency weighting filter processing part 39 has a structure as shown in Fig. 6A. The linear prediction coefficients a are supplied to a frequency weighting filter coefficient calculation part 39A in which coefficients γiai of a filter having a transfer characteristic A(z/γ) are calculated. A frequency weighting filter 39B calculates coefficients of a filter having a transfer characteristic Hw(z) = A(z)/A(z/γ) based on the linear prediction coefficients a and the frequency weighted coefficients γiai, and at the same time the phase-aligned speech Sp(t) is passed through the filter of this transfer characteristic to obtain a signal S'w(t).

A zero-input response calculation part 39C uses as an initial value a synthesized speech

obtained as the output of the all-pole filter 18A (see Fig. 1) of a transfer characteristic 1/A(z/γ) in the previous frame, and outputs an initial response when the all-pole filter 18A is excited by a zero input.

A target signal calculation part 39D subtracts the output of the zero input response calculation part 39C from the output signal S'w(t) of the frequency weighting filter 39B to obtain a frequency weighted signal Sw(t). On the other hand, the output γiai of the frequency weighting filter coefficient processing part 39A is supplied to an impulse response calculation part 40 in Fig. 6, in which an impulse response f(t) of a filter having the transfer characteristic 1/A(z/γ) is calculated.

A correlation calculation part 41 calculates for each pulse position ti a cross-correlation ψ(i) between the impulse response f(t-ti) and the frequency-weighted signal Sw(t) as follows:

where i = 1, 2 ..., np. np is the number of pulses in the frame, and N is the number of samples in the frame.

Another correlation calculation part 42 calculates a covariance φ(i, j) of the impulse response for a set of impulse positions ti, tj as follows:

A pulse magnitude calculation part 43 obtains pulse magnitudes mi of ψ(t) and φ(i, j) by solving the following simultaneous equations which equivalently minimize a mean square error between a synthesized speech waveform obtained by exciting the all-pole filter 18 with the pulse train thus determined and the phase-aligned speech waveform Sp(t).

The pulse sizes m are quantized by the quantizer 9 in Fig. 1 for each frame. This is carried out by, for example, a scalar quantization method or a vector quantization method. In the case of using the vector quantization technique, a vector (a magnitude pattern) using respective pulse sizes m as its elements is compared with a plurality of predetermined standard pulse size patterns and quantized to that one of them which minimizes the distance between the patterns. A measure of the distance between the magnitude patterns substantially corresponds to a mean square error of the speech waveform Sp'(t) synthesized without using the zero filter from the standard pulse size pattern selected in the quantizer 9 and the phase-aligned input speech waveform Sp(t). For example, denoting the magnitude pattern vector obtained by solving equation (11) by m = (m₁, m₂..., mnp) and If standard pattern vectors stored as a table in the quantizer 9 are denoted by mci (i = 1, 2 ..., Nc), then the mean square error is given by the following equation:

d(m, mc) = (m - mci)t&Phi;(m, mci) ... (12)

where t denotes the transposition of a matrix and Φ is a matrix using as its elements the auto-covariance Φ(i, j) of the impulse response. In this case, the quantized value m of the above-mentioned magnitude pattern is expressed by the following equation as a standard pattern which minimizes the mean square error d(m, mc) in equation (12) in the aforementioned plurality of standard pattern vectors mci.

The zero filter 10 serves to supply an input pulse train having a feature of the phase-matched prediction residual waveform, and the coefficients of this filter are generated by a zero filter coefficient calculation part 11. Fig. 7A shows an example of the phase-matched prediction residual waveform ep(t) and Fig. 7B shows an example of an impulse response waveform of the zero filter 10 for the input pulse thereto. The phase-matched prediction residual ep(t) has a flat spectral envelope characteristic and a phase close to zero and is thus pulse-like and has a large size at pulse positions ti, ti-1, ... but is relatively small at other positions. The waveform is substantially symmetrical with respect to each pulse position or each midpoint between adjacent pulse positions. In many cases, the magnitude at the center is relatively larger than at other positions (except the pulse positions) as seen from Fig. 7A, and this tendency increases particularly for a speech with a long pitch frequency. The null filter 10 is set so that its impulse response takes values at q consecutive sampling points on both sides of the pulse position t and at r consecutive sampling points on both sides of the center between the adjacent pulse positions ti and ti+1, as shown in Fig. 7B. In this case, the transfer characteristic of the null filter 10 is as follows:

In the zero filter coefficient calculation part 11, for a pulse train of given pulse positions and pulse sizes, filter coefficients vk are determined so that a frequency-weighted mean square error between the synthesized speech waveform Sp'(t) and the phase-aligned input speech waveform Sp(t) can be minimal. Fig. 8 shows the structure of the filter coefficient calculation part 11. A frequency weighting filter processing part 44 and an impulse response calculation part 45 are identical in structure to the frequency weighting filter processing part 39 and the impulse response calculation part 40 in Fig. 6, respectively. An adder 46 adds the output impulse response f(t) of the impulse response calculation part 45 according to the following equation:

where l = q + r + 1.

A correlation calculation part 47 calculates the cross-covariance φ(i) between the signals Sw(t) and ui(t), and another correlation calculation part 48 calculates the auto-covariance φ(i, j) between the signals ui(t) and uj(t). A filter coefficient calculation part 49 calculates coefficients vi of the zero filter 10 from the above-mentioned cross-correlation φ(i) and the covariance φ(i, j) by solving the following simultaneous equations:

These solutions ultimately minimize a mean square error between a synthesized speech waveform, obtainable by exciting the all-pole filter 18 with the output signal of the zero filter 10, and the phase-aligned speech waveform Sp(t).

The filter coefficient vi is quantized by a quantizer 12 in Fig. 1. This is done using, for example, a scalar quantization technique or a vector quantization technique. In the case of using the vector quantization technique, a vector (a coefficient pattern) using the filter coefficients vi as its elements is compared with a plurality of predetermined standard coefficient patterns and quantized into a standard pattern that minimizes the distance between the patterns. When a measure substantially corresponding to the mean square error between the synthesized speech waveform Sp'(t) and the phase-aligned input speech waveform Sp(t) is used as the distance measure in the case of vector quantization of the pulse size by the aforementioned quantizer 9, the quantized value v of the filter coefficients is obtained by the following equation:

where v is a vector using as its elements coefficients v-q, v-q+1, ..., vq+2r+1 obtained by solving equation (16), and vci is a standard pattern vector of the filter coefficients. Further, Φ is a matrix using as its elements the covariance Φ(i, j) of the impulse response ui(t).

To summarize, in the voiced sound frame, the speech signal Sp'(t) is synthesized by exciting an all-pole filter having the speech spectral envelope characteristic with a quasi-periodic pulse train determined by pulse positions based on the phase-matched residual ep(t) and pulse sizes determined so that an error of the synthesized speech is minimal. Of the excitation parameters, the pulse sizes m and the coefficients v of the zero filter are set to optimal values which minimize the matching error between the synthesized speech waveform Sp'(t) and the phase-matched speech waveform Sp(t).

Next, the excitation in the unvoiced sound frame is described. In the unvoiced sound frame, a random pattern is used as an excitation signal as is the case in code excited linear predictive coding (Schroeder, et al., "Code excited linear prediction (CELP)", IEEE Int. On ASSP, pp. 937-940, 1985). In a random pattern generator part 13 in Fig. 1, a plurality of patterns each composed of a plurality of normal random numbers having a mean of 0 and a variance of 1 are stored. A gain calculation part 15 calculates, for each random pattern, a gain gi which makes the power of the synthesized speech Sp'(t) by the output random pattern and the power of the phase-matched speech Sp(t) equal, and a scalar-quantized gain i from a quantizer 16 is used to control an amplifier 14. A matching error between a synthesized speech waveform Sp'(t) obtained by applying each of the random patterns to the all-pole filter 18 and the phase-matched speech Sp'(t) is obtained by a waveform matching error calculation part 19. The errors thus obtained are decided by the error decision part 20, and the random pattern generator part 13 searches for an optimal random pattern which makes the waveform matching error minimum. In this embodiment, a frame is composed of three consecutive random patterns. This random pattern sequence is applied as the excitation signal via the amplifier 14 to the all-pole filter 18.

Following the above procedure, the speech signal is represented by the linear prediction coefficients ai and the voiced/unvoiced sound parameter VU. The voiced sound is represented by the pulse positions ti, the pulse sizes i and zero filter coefficients i, and the unvoiced sound is represented by the random number code pattern (number) ci and the gain i. These speech parameters are encoded by an encoding part 21 and then transmitted or stored. In a speech synthesizing part 21, the speech parameters are decoded by a decoder part 22. In the case of the voiced sound, a pulse train composed of the pulse positions ti and the pulse sizes i is generated in a pulse train generator part 23 and applied to a zero filter 24 to provide an excitation signal. In the case of the unvoiced sound, a random pattern is selectively generated by a random pattern generator part 25 using the random number code (signal) c and is supplied to an amplifier 26 which is controlled by the gain i and in which it is magnitude-controlled to produce an excitation signal. One of the excitation signals thus produced is selected by a switch 27 which is controlled by the voiced/unvoiced parameter VU and the excitation signal thus selected is applied to an all-pole filter 28 for excitation thereof which provides a synthesized speech at its output end. The filter coefficients of the zero filter 24 are controlled by i and the filter coefficients of the all-pole filter 28 are controlled by ai.

In a first modified form of the above embodiment, in the structure of Fig. 1, the pulse excitation source is used in common for voiced and unvoiced sounds. That is, the random pattern generator part 13, the amplifier 14, the gain calculation part 15, the quantizer 16 and the switch 17 are omitted, and the output of the zero filter 10 is directly applied to the all-pole filter 18. This slightly deteriorates the speech quality for a fricative consonant, but allows the structure for processing to be simplified and enables the amount of data to be processed to be reduced. Thus, the amount of hardware used can be small. In addition, since the voiced/unvoiced sound parameter does not need to be transmitted, the bit rate is reduced by 60 bits per second.

In a second modified form, the zero filter 10 is not included in the pulse excitation source in Fig. 1, that is, the zero filter 10, the zero filter coefficient calculation part 11 and the quantizer 12 are eliminated, and the output signal of the pulse train generator part 7 is supplied to the all-pole filter 18 via the switch 17. (The zero filter 24 is also eliminated accordingly.) In this method, the natural sound characteristic of the synthesized speech is slightly deteriorated for a speech of a male voice with a low pitch frequency, but the removal of the zero filter 10 reduces the amount of hardware used and the bit rate is reduced by 600 bits per second required for coding filter coefficients.

In a third modified form, the processing by the pulse magnitude calculation part 8 and the processing by the vector quantization part 9 in Fig. 1 are integrated to calculate a quantized value of the pulse magnitudes. Fig. 9 shows the structure of this modified form. A frequency weighting filter processing part 50, an impulse response calculation part 51, a correlation calculation part 52 and another correlation calculation part 53 are identical in structure to those in Fig. 6. In a pulse magnitude (vector) quantization part 54, for each pulse standard pattern mci (where i = 1, 2 ..., Nc), a mean square error between a speech waveform synthesized using the magnitude standard pattern and the phase-aligned input speech waveform Sp(t) is calculated, and a pulse magnitude standard pattern that minimizes the error is obtained. A distance calculation is performed using the following equation:

d = mcit&Phi;mci - 2mcit&psi;

where Φ is a matrix that uses the covariance Φ(i, j) of the impulse response f(t) as matrix elements, and Ψ is a column vector that has as its elements the cross-correlation Ψ(i) (where i = 1, 2, ..., np) of the impulse response and the output Sw(t) of the frequency weighting filter processing part 50.

The structures in Figs. 6 and 9 are almost identical in terms of the amount of data to be processed to obtain the optimum pulse size, but in Fig. 9 the processing for solving the simultaneous equations, which is included in the processing in Fig. 6, is not required and the processor is therefore simple. In Fig. 6, however, the maximum value of the pulse size can be scalar-quantized, while in Fig. 9 it is assumed that the vector quantization method is used.

It is also possible to calculate quantized values of coefficients by integrating the calculation of the coefficients vi of the zero filter 10 and the vector quantization by the quantizer 12 in the same way as mentioned above with reference to Fig. 9.

In a fourth modified form of the embodiment of Fig. 1, the pulse position generator part 6 is not provided and hence the processing shown in Fig. 4 is not included. Instead, all reference instants t'i provided by the phase alignment analyzer part 4 are used as pulse positions ti. This slightly increases the amount of information required to encode the pulse positions, but simplifies the design and speeds up the processing. Nevertheless, the throughput for increasing the quality of the synthesized speech by using the zero filter 10 can also be allocated to reducing the pulse position information at the expense of speech quality.

It is evident that in the embodiments of the speech analysis synthesis apparatus according to the present invention, the illustrated functional blocks thereof may be constituted by hardware and functions of some or all may be executed by means of a computer.

To estimate the effect of the speech analysis synthesis method according to the present invention, experiments were carried out using the following conditions. After sampling a speech in a 0 to 4 kHz band at a sampling frequency of 8 kHz, the speech signal is multiplied by a Hamming window of an analysis window of 30 ms in length, and linear predictive analysis is carried out by an autocorrelation method with the degree of analysis set to 12, thereby obtaining 12 prediction coefficients ai and the voiced/unvoiced sound parameter. The processing of the excitation parameter analyzer part 30 is carried out for each frame of 15 ms (120 speech samples) equal to half of the analysis window. The prediction coefficients are quantized by a differential multi-stage vector quantization method. A frequency-weighted cepstrum distance was used as a distance criterion in the vector quantization. When the bit rate is 4.8 kb/s, the number of bits per frame is 72, and the details are as follows: Parameter Number of bits/frame Prediction coefficients voiced/unvoiced Sound parameters Excitation source (for voiced sound) Pulse positions Pulse sizes Zero filter coefficients Excitation source (for unvoiced sound) Random pattern Gains

The constant J representing the permissible limit of fluctuations of the pulse frequency in the pulse source, the permissible maximum number Np of pulses per frame and the permissible minimum value Lmin of pulse intervals depend on the number of bits allocated for coding the pulse positions. For example, in the case of coding the pulse positions at a rate of 29 bits/frame, it is favorable that the difference between adjacent pulse intervals ΔT is equal to or less than 5 sampling points, the maximum number Np of pulses is equal to or less than 6 sampling points and the permissible minimum pulse interval Lmin is equal to or greater than 13 sampling points. A seventh-order filter (q = r = 1) was used as the zero filter 10. The random pattern vector ci is composed of 40 samples (5 ms) and is selected from 512 kinds of patterns (9 bits). The gain gi is scalar-quantized using 6 bits including a sign bit.

The speech encoded using the above conditions sounds much more natural than the speech encoded by the conventional vocoder, and its quality is close to that of the original speech. Furthermore, the dependence of the speech quality on the speaker is lower in the present invention than in the case of the known vocoder. It has been proved that the quality of the encoded speech is significantly higher than in the case of the conventional multi-pulse predictive coding and the code-excited predictive coding. A spectral envelope error of a speech encoded at 4.8 kb/s is about 1 dB. An encoding delay of this invention is 45 ms, which is equal to or shorter than that of the conventional low-bit-rate speech coding schemes.

A short Japanese sentence uttered by two men and two women was speech analyzed using substantially the same conditions as mentioned above to obtain the excitation parameters, the prediction coefficients and the voiced/unvoiced parameter VU, which were then used to synthesize a speech. An opinion test for subjective quality evaluation of the synthesized speech was carried out by 30 persons. In Fig. 10, the results of the test are shown in comparison with those in the cases of other coding methods. The abscissa shows MOS (mean opinion score) and ORG of the original speech. PCM4 to PCM8 represent synthesized speeches by 4 to 8 bit log-PCM coding methods, and EQ denotes a phase-aligned speech. The test results show that the coding by the present invention is performed at a low bit rate of 4.8 kb/s, but a high quality synthesized speech which is equal in quality to the speech synthesized by the 8-bit Log-PCM coding.

According to the present invention, by expressing the excitation signal for a voiced sound as a quasi-periodic pulse train, the repeatability of the speech waveform information is higher than in the conventional vocoder, and the excitation signal can be expressed with a smaller amount of information than in the conventional multi-pulse predictive coding. Moreover, since an error between the input speech waveform and the phase-matched speech waveform is used as the criterion for estimating the parameters of the excitation signal from the input speech, the present invention increases the agreement between the synthesized speech waveform and the input speech waveform compared with the prior art which uses an error between the input speech waveform itself and the synthesized speech. Thus, the invention allows accurate estimation of the excitation parameters. In addition, the null filter produces the effect of reproducing subtle spectral characteristics of the original speech, whereby the synthesized speech will sound more natural.

It will be apparent that many modifications and variations may be made without departing from the scope of the novel concept of the invention as claimed in the appended claims.

Claims (9)

1. A method of analyzing speech for generating an excitation signal for exciting a linear filter representing a speech spectral envelope characteristic, comprising: a step of determining positions of pulses and generating a train of pulses at the determined positions to express the excitation signal, a step of determining parameters representing the excitation signal so that an error between a phase-aligned speech waveform after phase-aligning an input speech and a synthesized speech waveform obtainable by exciting the linear filter with the pulse train is minimized, a step of generating a phase-aligned prediction residual of the input speech waveform, characterized in that the step of determining the pulse positions and generating the pulse sequence expressing the excitation signal includes: a step of determining reference time points where levels of the phase-adjusted prediction residual exceed a predetermined threshold value, and a step of determining pulse positions of a quasi-periodic pulse train as the pulse positions of the pulse train on the basis of the reference time points such that a fluctuation of successive time intervals of the pulse positions is within a limited range. 2. The method according to claim 1, comprising a step of determining coefficients of a zero filter (10) characterizing a fine spectral structure of the speech so that an error between the phase-aligned speech waveform and a synthesized speech waveform obtainable by exciting the linear filter (18) with the output signal of the zero filter becomes minimal, the coefficients of the zero filter being used as one of the parameters representing the excitation signal. 3. A method according to claim 1 or 2, wherein the excitation signal for a voiced sound is used and a random sequence selected from a plurality of random patterns is used as an excitation signal for an unvoiced sound, and which includes a step of determining parameters representing the excitation signal for the unvoiced sound so that an error between the phase-aligned speech waveform and a synthesized speech waveform obtainable by exciting the linear filter with the random patterns becomes minimal. 4. A method according to claim 1 or 2, wherein the parameters representing the excitation signal include a parameter which determines the size of each individual pulse of the pulse sequence wherein the magnitude parameter is determined such that an error between the phase-aligned speech waveform and a synthesized speech waveform obtainable by exciting the linear filter with the pulse train is minimized. 5. Speech analysis device comprising: linear predictive analysis means (2) for performing a linear predictive analysis of an input speech signal (s(t)) for each analysis window of a fixed length to obtain prediction coefficients (ai), inverse filter means (31) controlled by the prediction coefficients for deriving a prediction residual (e(t)) from the input speech signal (s(t)), Speech phase adjustment filter means (5) for converting the phase of the input speech signal to a zero phase to obtain a phase-adjusted speech signal (Sp(t)), Prediction residual phase adjustment filter means (37) for converting the phase of the prediction residual (e(t)) to a zero phase to obtain a phase-adjusted prediction residual signal ep(t), Means (4, 6, 7) for determining pulse positions and generating a sequence of pulses at the positions as an excitation signal, All-pole filter means (18) controlled by the prediction coefficients and excited by the pulse sequence for generating a synthesized speech, and Pulse magnitude calculation means (8) for determining magnitudes of the pulse train so as to minimize an error between a waveform of synthesized speech obtainable by exciting the all-pole filter means with the pulse train and a waveform of the phase-aligned speech, parameters including the pulse positions and the magnitudes of the pulses being output from the speech analyzing device, characterized in that the means (4, 6, 7) for determining the pulse positions and generating the pulse sequence contain: Reference time point generator means (4, 38) for detecting pulse sizes greater than a predetermined threshold value in the phase-adjusted prediction residual signal and for outputting the positions of the pulses as reference times, and Pulse position generator means (6) for determining, on the basis of the reference times, the positions of the pulses having a pitch frequency with a limited fluctuation range. 6. The device of claim 5, further comprising: null filter means (10) supplied with the pulse train for supplying the pulse train having characteristics of the waveform of the phase-aligned prediction residual signal and for supplying its output to the all-pole filter means (18) as the excitation signal, and null filter coefficient calculating means (11) for determining coefficients of the null filter means such that an error between a waveform of a synthesized speech obtained by exciting the all-pole filter means with the output of the null filter means and a waveform of the phase-aligned speech becomes minimal. 7. Apparatus according to claim 5 or 6, wherein the linear predictive analysis means (2) include means for determining whether the input speech signal is in an analysis window of a fixed length is voiced or unvoiced, and for outputting a voiced/unvoiced decision signal (VU), the apparatus further comprising random pattern generator means (13) for generating a random pattern which minimizes an error between a waveform of a synthesized speech obtained by exciting the all-pole filter means (18) with one of a plurality of predetermined random patterns and a waveform of the phase-aligned speech in a window during which the decision signal is unvoiced. 8. Device according to claim 5 or 6, wherein the means (4, 6, 7) for generating the pulse sequence contain vector quantization means (9) for vector quantizing the magnitude values of the pulses determined by the pulse magnitude calculation means (8), whereby the pulse sequence has the quantized magnitude values. 9. A speech synthesis apparatus that synthesizes a speech in response to parameters representing an excitation signal output from a speech analysis apparatus according to claim 5 or 6, comprising: Pulse train generator means (23) for generating a pulse train based on the parameters, zero filter means (24) excited by the pulse train under the control of zero filter coefficients supplied thereto as one of the parameters, for creating the pulse train with a spectral characteristic of the speech, and All-pole filter means (28) excited by the output of the null filter means under the control of prediction coefficients representing a speech spectral envelope characteristic, for synthesizing a speech waveform.