EP0402947B1

EP0402947B1 - Arrangement and method for encoding speech signal using regular pulse excitation scheme

Info

Publication number: EP0402947B1
Application number: EP19900111360
Authority: EP
Inventors: Yoshihiro C/O Nec Corporation Unno
Original assignee: NEC Corp
Current assignee: NEC Corp
Priority date: 1989-06-14
Filing date: 1990-06-15
Publication date: 1997-11-26
Anticipated expiration: 2010-06-15
Also published as: DE69031749T2; EP0402947A2; DE69031749D1; EP0402947A3

Description

The present invention relates generally to an arrangement and method for encoding a discrete-time speech signal using a regular pulse excitation scheme and more specifically to such an arrangement and method for encoding a speech signal at a low bit rate less than 16k-bit per second.
In order to encode a speech signal with the limited number of calculations at a low bit rate (less than approximately 16k-bit per second), it is a known practice to model the characteristics of a human's vocal tract using a digital filter and further to exhibit excitation signals by combining regular pulse sequences. Such a coding scheme is known as a Regular Pulse Excitation - Long Term Prediction - Linear Predictive Coder (hereinlater referred to as RPE-LTP), which has been proposed in the CEPT/CCH/GSM Recommendation 06.10 entitled "GSM Full Rate Speech Transcoding" published by Conference of European Postal and Telecommunications Administrations, September 19, 1988 (hereinafter referred to as Paper 1).
Before describing the present invention, the regular pulse excitation coding scheme disclosed in Paper 1 will be described with reference to Fig. 1.
In Fig. 1, an a/d (analog-to-digital) converted speech signal is applied via an input terminal 10 to a pre-processing circuit 12 on a frame by frame basis. The speech frame applied to the circuit 12 is pre-processed to produce an offset-free signal, which signal is then subjected to a first order pre-emphasis filter. An original speech signal has been sampled at a rate of 8 kHz. Since the frame length is 20 ms in this prior art, the one frame consists of 160 signal samples. The 160 samples thus obtained are applied to a short term LPC (Linear Predictive Coding) analysis circuit 14 and also to a short term analysis filter 16. The 160 samples, applied to the short term LPC analysis circuit 14, are analyzed to determine 8 orders of reflection coefficients which represent a spectrum envelope of each frame. The LPC short term analysis circuit 14 further transforms or encodes the reflection coefficients to log area ratios (LAR), which are applied to the short term analysis circuit 16 and a multiplexor 30. The short term analysis circuit 16 decodes the LAR into the reflection coefficients and obtains 160 samples of short term residual signals. In the above, the term "short term analysis" has the same meaning as the spectrum envelope analysis. The short term residual signal, outputted from the filter 16, is applied to a subtractor 18 and a long term analysis circuit 22.
For the following operations, the long term analysis circuit 22 divides the speech frame into 4 sub-frames (5 ms) each of which consists of 40 samples forming the short term residual signal. Each sub-frame is processed blockwise by the subsequent function blocks.
The long term analysis circuit 22 produces a long term prediction (LTP) lag and an LTP gain on the basis of the two signals: the short term residual samples applied from the circuit 16 and an output sequence from an adder 26. The term "long term analysis" has the same meaning as pitch analysis, and the LTP lag and the LTP gain respectively correspond to a pitch period and a pitch gain.
The subtractor 18 outputs a block of 40 long term residual signal samples by subtracting the output of a long term analysis filter 20 from the short term residual signal applied from the filter 16. An excitation pulse calculating circuit 24, using the long term residual signal samples from the subtractor 18, obtains an RPE grid of an excitation pulse sequence and an amplitude sequence of an excitation pulse series, which are encoded and fed to the multiplexor 30 and also to an excitation pulse generator 28. In connection with an RPE grid, reference should be made to Paper 1.
The position of j-th excitation pulse (m_j) within a sub-frame is given by the following equation. $m_{j} = P·j + q (0 ≤ j ≤ N/p - 1 and 0 ≤ q ≤ p)$
where p denotes a predetermined pulse interval, q an RPE grid, and N the number of samples within one sub-frame. By expressing the output sequence of the subtractor 18 as x_j(i), the RPE grid q of the excitation pulse sequence is obtained from the following equation.
where max(q) indicates the maximum value of the right term when changing the value of q. The amplitude of the excitation pulse sequence can be determined by quantizing x_j(m_j).
The excitation pulse generator 28 decodes the signal applied from the circuit 24 to determine an excitation pulse, which is fed to the adder 26. The adder 26 adds the excitation pulse from the circuit 28 and the output sequence of the long term analysis filter 20, and applies the resultant sum to the filter 20 as well as the analysis circuit 22. The long term analysis filter 20, utilizing the LTP lag and the LTP gain, both applied from the circuit 22, filters the output sequence of the adder 26. The output sequence of the filter 20 is fed back to the adder 26 and also applied to the subtractor 18.
The multiplexor 30 combines the encoded outputs of the blocks 14, 22 and 24, and applies the result to a transmission line coupled to an output terminal 32.
However, the above-mentioned prior art has encountered the difficulty of low quality of the reconstructed or reproduced speech. This is because the amplitude of each excitation pulse is determined on the basis of the short term residual signal applied to the subtractor 18. In other words, according to the prior art, the long term residual signal outputted from the subtractor 18 is shifted by an RPE grid and then every predetermined number of samples are quantized.
Furthermore, the aforesaid prior art has encountered another problem in that the reproduced speech is degraded by quantizing distortion. This results from the fact that the number of quantizing bits is insufficient at a bit rate in the order of 13k bps.
EP-A-0 374 941 (document according to Article 54(3) EPC) discloses a communication system for improving the speech quality by calculating the excitation multi-pulses by means of an encoder for encoding a sequence of digital speech signals classified into a voiced sound and an unvoiced sound into a sequence of output signals by the use of a spectrum parameter and pitch parameters at every frame. A judging circuit judges whether the digital speech signals are classified into the voiced sound or the unvoiced sound in order to produce a judged signal representative of a result of judging. A processing unit processes the digital speech signals in accordance with the judged signal to selectively produce a first set of primary sound source signals and secondary sound source signals. This first set is produced when the judged signal represents the voiced sound and are representative of locations and amplitudes of a first set of excitation multi-pulses calculated at every frame. The second set of secondary sound source signals are produced when the judged signal represents the unvoiced sound and are representative of the amplitudes of a second set of excitation multi-pulses each of which is located at intervals of a preselected number of the samples.
The document ICASSP 87, Dallas, Texas, 6th - 9th April 1987, vol. 2, pages 968-971, IEEE, New York, US; A. Fukui et al: "Implementation of a multi-pulse speech codec with pitch prediction on a single chip floating-point signal processor" discloses a system of multi-pulse codec with pitch prediction which divides a discrete-time speech signal and extracts a plurality of parameters from the divided speech signals. The parameters are used to generate a signal and for generating an impulse response function signal which, in turn, are used for generating an autocorrelation function signal and a cross-correlation function signal.
It is an object of the present invention to provide an arrangement for encoding a discrete-time speech signal using a regular pulse excitation scheme.
It is an object of the present invention to provide an arrangement for encoding a discrete-time speech signal at a low bit rate less than 16k-bit per second through the use of a regular pulse excitation scheme.
Another object of the present invention is to provide a method for encoding a discrete-time speech signal at a low bit rate less than 16k-bit per second using a regular pulse excitation scheme.
These objects are solved with the features of the claims.
In brief, a binary adder is comprised of a pre-processing circuit provided to receive a discrete-time speech signal which are then divided into a plurality of frames. A parameter extracting circuit is coupled to the pre-processing circuit and extracts a plurality of parameters therefrom. A impulse response calculating circuit is coupled to receive the plurality of parameters from the parameter extracting circuit, and generates an impulse response function signal using the plurality of parameters. An autocorrelation function circuit is coupled to receive the impulse response signal and generates an autocorrelation function signal using the signal applied. A cross-correlation function circuit generates a cross-correlation function signal using the discrete-time speech signal and the autocorrelation function signal. A grid signal generator receive the output of the cross-correlation function calculating circuit, and outputs a grid signal indicative of a location of a first excitation pulse within one frame. A pulse amplitude calculating circuit receives the autocorrelation function signal, the cross-correlation function signal and the grid signal, and determines an amplitude sequence of excitation pulses within one frame.
one aspect of this invention takes the form of an arrangement for encoding a speech signal using a regular pulse excitation scheme, as set out in the appended claims.
Another aspect of this invention takes the form of a method for encoding a speech signal using a regular pulse excitation scheme, as set out in the appended claims.
The features and advantages of the present invention will become more clearly appreciated from the following description taken in conjunction with the accompanying drawings in which like elements are denoted by like reference numerals and in which:

Fig. 1 is a block diagram illustrating a known RPE scheme, the drawing having been referred to in the opening paragraphs of this specification;
Fig. 2 is a block diagram showing a first embodiment of this invention;
Figs. 3 and 4 each is a block diagram showing in detail a block in Fig. 2; and
Fig. 5 is a block diagram showing a second embodiment of this invention.

The present invention is characterized by algorithms for calculating an amplitude of each of the excitation pulses. It should be noted that the location of the excitation pulse can be determined in accordance with the prior art disclosed in Paper 1. The above mentioned algorithms will be discussed below.
According to a so-called RPE coding scheme, the location of a j-th excitation pulse within a frame can be specified by equation (1). For the convenience of description, equation (1) is again shown as equation (3). $m_{j} = p·j + q (0 ≤ j ≤ N/p -1 and 0 ≤ q ≤ p)$
Algorithm of obtaining the RPE grid q will be described later.
An excitation pulse sequence d(n) can be represented by
where n denotes a given time within one frame, g_i an amplitude of an excitation pulse located at a position m_i, δ(n,m_i) the Kronecker's delta function which assumes 1 in the case of n = m_i and 0 in the case of n ≠ m_i, and K represents the number of pulses within one frame.
Fig. 3 shows a synthesis filter 122 which comprises two digital filters 310 and 320 coupled in series. The filter 310 includes an adder 322, a coefficient weighting circuit 324 and a delay 326. Similarly, the filter 320 includes an adder 328, a coefficient weighting circuit 330 and a delay 332. The synthesis filter 122 forms part of the arrangement shown in Fig. 2, and will again be referred to later. Consequently, the detail description of Fig. 3 will be postponed.
The filter 310 is a long term prediction filter whose output represents a pitch structure, while the filter 320 is a short term prediction filter whose output represents spectrum envelope characteristics. For simplifying the description, it will be assumed that the filter 310 is of a first order type. The synthesis filter 122 is supplied with the excitation pulse series and outputs a reconstructed signal sequence x'(n) in accordance with the following equation:
where β denotes an LTP gain representative of tap coefficients of the long term filter 310, Md a LTP lag indicative of a pitch period of an incoming speech signal. Further, in equation (5), x_d(n) denotes an output signal of the filter 310, Np a prediction order of the short term prediction filter 320, and a_i (1≤ i ≤Np) a prediction coefficient of the filter 320 (a_i corresponds to LAR in Fig. 3). β and Md can be obtained in accordance with the prior art techniques disclosed in Paper 1. As an alternative, β and Md can be determined by a peak amplitude of the autocorrelation function sequence of an input speech signal and the position of said peak. The algorithms via which this can be achieved have been disclosed in the document entitled "Adaptive predictive coding of speech signals" by B.S. Atal et al., pages 1973 to 1986, The Bell System Technical Journal, October 1970 (referred to as Paper 2).
By defining the impulse response of the synthesis filter 122 as h(i) (0 ≤ i ≤ M-1 (M is the number of continuous samples)), the reconstructed signal x'(n) is given by: $x'(n) = d(n) * h(n)$
where the symbol * denotes convolution integration. Further, the square error J in weighting between the input speech signal x(n) and the reproduced signal x'(n) within one frame, can be represented by:
where N denotes the number of samples within one frame and w(n) a weighting function. The weighting function w(n) implements weighting on a frequency axis, and the Z transform W(Z) thereof is given by:
where a_i represents a prediction parameter of the synthesis filter 122 and r is a constant (0≤ r ≤1) which determines the frequency characteristics of W(Z). In more detail, in the event of r=1 then W(Z)=1. This means that the frequency characteristics is flat. On the other hand, when r=0 then W(Z) represents an inverse frequency characteristics of the synthesis filter 122. It follows that the value of r is able to change the characteristics of W(Z). The reason why W(Z) is determined depending upon the frequency characteristics of the synthesis filter 122 as shown in equation (8), stems from the fact that an unaudible masking effect is utilized. In more detail, at a portion where the power of the spectrum of the input speech signal is large (for example in the vicinity of a formant), even if the difference or error between the spectrums of input and reconstructed signals is somewhat large, such error does not affect the hearing sense of the ears.
Algorithms for calculating an excitation pulse series which minimizes the weighted square error J shown in equation (7), will be discussed in the followings. Equation (7) is rewritten as follows.
where the term x'(n) * w(n) can be modified according to the following equation. Thus by putting $x_{w}'(n) = x'(n) * w(n)$
and by performing Z conversion on both sides of equation (10), we obtain: $X_{w}'(Z) = X'(Z)·W(Z)$
Further, X'(Z) can be expressed as follows: $X'(Z) = H(Z)·D(Z)$
where D(Z) represents the Z conversion of the excitation pulse series given by equation (4), and H(Z) the Z conversion value of the impulse response of the synthesis filter 122. Substituting equation (12) into equation (11) gives: ${X'}_{w} (Z) = H(Z)·D(Z)·W(Z)$
By setting Hw(Z)=H(Z)·W(Z) and then implementing an inverse Z conversion on equation (13), we obtain ${x'}_{w} {(n) = d(n) * h}_{w} (n)$
where h_w(n) denotes an inverse Z conversion value of H_w(Z) and indicates the impulse response of a cascade coupled filter comprising a synthesis filter and a weighting circuit. By substituting equation (4) into equation (14), we obtain
where K represents the number of pulses within one frame. By substituting equations (10) and (15) into equation (9), we obtain
Thus, equation (7) can be rewritten into equation (16).
The following equation can be obtained by partially differentiating equation (16) with g_k and then setting it to zero, where g_k is an amplitude of the excitation pulse for minimizing equation (16). $g_{k} {(m}_{k}) = \frac{φ_{xh} {(-m}_{k} {) - g}_{1} {·φ}_{hh} {(m}_{1} {,m}_{k} {) - ··· - g}_{k-1} {·φ}_{hh} {(m}_{k-1} {,m}_{k})}{φ_{hh} {(m}_{k} {,m}_{k})}$
where φ_xh(·) represents a cross-correlation function sequence computed from x_w(n) and h_w(n), and φ_hh(·) represents an autocorrelation function sequence of hw(n). These two sequences are represented by the following equations (18) and (19). φ_hh(·) is referred to as a covariance function in the art of speech signal processing.
As will be understood from equation (17), the amplitude g_k of each of the excitation pulses is a function of the location m_k of the corresponding excitation pulse. This means that the most desirable amplitude g_k at a given pulse position m_k can be computed. If an incoming speech signal sequence is assumed stationary, then the covariance function φ_hh(m_i,m_k) can be represented by the following equation (20). $φ_{hh} {(m}_{i} {,m}_{k} {) = R}_{hh} (|m_{i} {- m}_{k}|)$
This equation indicates that under the above-mentioned assumption, φ_hh(m_i,m_k) is equal to an autocorrelation function R_hh(·) which depends on a delay $|m_{i} {- m}_{k}|$
.
R_hh( $|m_{i} {- m}_{k}|$
) in equation (20) can be represented as follows:
Consequently, equation (17) can be modified using equation (21) as follows: $g_{k} {(m}_{k}) = \frac{φ_{xh} {(-m}_{k} {) - g}_{1} {·R}_{hh} (|m_{1} {- m}_{k}|) -}{R_{hh} (0)} \frac{{- g}_{k-1} {·R}_{hh} (|m_{k-1} {- m}_{k}|)}{R_{hh} (0)}$
The value of RPE grid q is calculated using the cross-correlation function obtained by equation (18). That is to say, the RPE grid q can be determined so as to satisfy the following equation.
where max(q) indicates the maximum value of the right term when changing the value of q. The value that an RPE grid q can assume is 0, 1, 2, 3 in the prior art disclosed in Paper 1 merely by way of example.
According to the present invention, an amplitude sequence of the excitation signal can be precisely obtained using equation (22), and hence a high quality reproduced voice can be realized.
A first embodiment of this invention will be discussed with reference to Figs. 2 to 4.
As previously mentioned in connection with Fig. 1, an a/d (analog-to-digital) converted speech signal is applied via an input terminal 110 to a pre-processing circuit 112 on a frame by frame basis. The pre-processing circuit 112 can be configured in the same manner as the circuit 12 of Fig. 1. The speech frame applied to the circuit 112 is pre-processed to produce an offset-free signal, which is then subjected to a first order pre-emphasis filter. An original speech signal to be applied to the input terminal 110, has been sampled at a predetermined rate such as 8 kHz. In the event that the frame length is 20 ms as in the prior art, merely by way of example, the one frame consists of 160 signal samples. The samples thus obtained are applied to a short term LPC (Linear Predictive Coding) analysis circuit 114 and also to a long term (pitch) analysis filter 116.
The one frame samples, applied to the short term LPC analysis circuit 114, are analyzed to determine predetermined orders of reflection coefficients (LAR(i)) (i=1···8) in the same manner as disclosed in Paper 1. The reflection coefficients represent a spectrum envelope of each frame. An LAR coding circuit 118 is supplied with the LAR(i)s and transforms or encodes them into log area ratios (coded-LAR(i)) based on predetermined quantizing levels (quantizing bits), and then applies them to a multiplexor 300. Further, the LAR coding circuit 118 decodes the coded-LAR(i)s, applies the decoded LAR'(i) to an impulse response calculating circuit 120 as well as a synthesis filter 122.
The long term analysis circuit 116 receives the one frame samples from the pre-processing circuit 112 to calculate LTP lag Md and LTP gain β along with the algorithms as disclosed in the above-mentioned Paper 2. The Md, β are fed to a long term (pitch) coding circuit 124, which encodes the Md, β and applies the coded-Md and coded-β to the multiplexor 300. Further, the long term coding circuit 124 decodes the coded-Md and the coded-β into Md' and β', respectively. The decoded LTP lag (Md') and the decoded LTP gain (β') are applied to the impulse response calculating circuit 120 and also to the synthesis filter 122.
As shown in Fig. 4, the impulse response calculating circuit 120 comprises an impulse generator 400, a long term prediction (LTP) filter 402 and a short term prediction (STP) filter 404, which are coupled in series. The LTP filter 402 includes an adder 406, a coefficient weighting circuit 408 and a delay circuit 410. Similarly, the STP filter 404 includes an adder 412, a coefficient weighting circuit 414 and a delay circuit 416. The operation of each of the filters 402 and 404 are known to those in the art, and hence the detail descriptions thereof will be omitted. The decoded Md' and β' are applied to the coefficient weighting circuit 408, while the decoded LAR'(i) to the coefficient weighting circuit 414.
The impulse response calculating circuit 120 determines an impulse response of a predetermined number of samples and applies the output h_w(n) to an autocorrelation function calculating circuit 126 and a cross-correlation function calculating circuit 128.
The circuit 126 calculates an autocorrelation function R_hh( $|m_{i} {- m}_{k}|$
) according to equation (21), and applies the result to a pulse amplitude calculating circuit 132.
A subtractor 134, coupled to the pre-processing circuit 112 and the synthesis filter 122, subtracts the output sequence of the filter 122 from the speech signal sequence x(n), and applies the resultant difference to a weighting circuit 136. The synthesis filter 122 has already stored one frame of response signal sequence, which is obtained by using an excitation pulse one frame before the present frame as an excitation signal and thereafter delayed to the present frame by making the excitation signal zero. This is based on a consideration that if it is assumed that the effective sample number of the impulse response of the synthesis filter in question is at most about two frames, the speech signal sequence of the present frame can be expressed by the sum of a signal sequence obtained by delaying the output signal of the synthesis filter driven by an excitation pulse one frame before to the present frame by making the excitation signal zero, and by the output signal sequence of the synthesis filter driven by the excitation pulse sequence of the present frame.
The weighting circuit 136 is supplied with the parameter LAR'(i) from the LAR coding circuit 118, and calculates the weighting function w(n) in a manner that the Z conversion value thereof satisfies equation (8). This calculation can be implemented through the use of another frequency weighting scheme. The weighting circuit 136 performs a convolution integration of the difference from the subtractor 134 and the function w(n), and applies the output thereof x_w(n) to the cross-correlation function circuit 128. This circuit 128 is further supplied with the impulse response hw(n), and calculates the cross-correlation function φ_xh(-m_k) (where 1≤ m_k ≤N) which is applied to a RPE grid selector 130 and also to the pulse amplitude calculating circuit 132.
The grid selector 130 determines or selects a grid q, using the cross-correlation function φ_xh(-m_k), according to equation (23) and applies the selected grid to the pulse amplitude calculating circuit 132. The circuit 132 is synchronously supplied with the above-mentioned three outputs (viz., the autocorrelation function R_hh( $|m_{i} {- m}_{k}|$
), the cross-correlation function φ_xh(-m_k) and the selected grid q), and determines an amplitude of each of the excitation pulses within one frame. In other words, the circuit 132 determines a so-called amplitude sequence of the excitation pulses in one frame.
A pulse coding circuit 137 receives the output sequence of the circuit 132 and encodes the selected grid q and the amplitude sequence g_k of the excitation pulses using normalizing coefficients, and applies the encoded information to the multiplexor 300. The normalizing coefficients are also encoded within the pulse coding circuit 137 and applied to the multiplexor 300. The circuit 137 further decodes the encoded data (viz., the grid and the amplitude sequence and the normalizing coefficients) to apply them to a pulse sequence generator 138. The decoded grid and the decoded amplitude sequence are respectively denoted by q' and g_k'. The operation of the pulse coding circuit 137 has been disclosed in the above-mentioned Paper 1.
The pulse sequence generator 138 outputs an excitation pulse sequence of one frame using g_k' and m_k', which pulse sequence has an amplitude g_k' at a position m_k'.
The synthesis filter 122 receives the excitation pulse sequence, and also receives the coefficients LAR'(i) and the pitch information (Md' and β') from the circuits 118 and 124, respectively. It should be noted that the synthesis filter 122 converts LAR'(i) into a prediction parameter a_i (1≤ i ≤Np) by means of a well known method. The filter 122 adds the excitation signal applied thereto and one frame of 0 sequence together with to determine a response signal sequence x(n) for the two frame signal. The sequence x'(n) can be represented by:
This equation is identical to equation (5). The excitation signal d(n) represents the output pulse signal generated by the pulse generating circuit 138 when 1≤ n ≤N, while representing a series of all zeros in the case of (N + 1)≤ n ≤2N. The subtractor 134 receives x'(n) obtained using equation (24) (wherein N + 1≤ n ≤2N).
The multiplexor 300 combines the outputs of the circuits 137, 118 and 124, which are applied to a transmission line via an output terminal 302.
A second embodiment of this invention will be discussed with reference to Figs. 3 to 5. The arrangement of Fig. 5 differs from that of Fig. 2 in that the former arrangement further includes a switch 500, a decision circuit 502, a gate 504 and a section 506. This section 506 is arranged in exactly the same manner as the arrangement of a section 508, although the functions of the two sections 506 and 508 are slightly different. For the convenience of description, each of the blocks 120', 126', 128', 130', 132' and 136' in the section 506 bears the same reference numeral as the counterpart in the section 508 but has a prime for the purposes of differentiation. The section 508 operates in the same manner as described above and hence further descriptions thereof will be omitted for simplicity. Similarly, in the case where the blocks included in the section 506 operates in the same manner as their counterparts in the section 508, the operations thereof may not be described for simplicity.
The impulse response calculating circuit 120' in the section 506 receives the decoded LAR'(i) at the coefficient weighting circuit 414 (Fig. 4), and determines an impulse response of a predetermined number of samples and applies the output h_w'(n) to the autocorrelation function calculating circuit 126' as well as the cross-correlation function calculating circuit 128'. This means that the circuit 120' utilizes only the short term prediction filter 404. It should be noted that as shown in Fig. 5 the line provided for the pitch information (Md' and β') is not coupled to the block 120' for disabling the long term prediction filter 402.
The autocorrelation function calculating circuit 126' calculates an autocorrelation function R_hh'( $|m_{i} {- m}_{k}|$
) according to equation (21), and applies the result to the pulse amplitude calculating circuit 132'. The weighting circuit 136' operates in the same manner as the counterpart 136, and applies the output thereof x_w(n) to the cross-correlation function calculating circuit 128'. This circuit 138' is further supplied with the impulse response hw'(n), and calculates the cross-correlation function φ_xh'(-m_k) (where 1≤ m_k ≤N) which is applied to the RPE grid selector 130' and also to the pulse amplitude calculating circuit 132'.
The grid selector 130' determines or selects a grid q', using the cross-correlation function φ_xh'(-m_k), according to equation (23) and applies the selected grid q' to the pulse amplitude calculating circuit 132'.
The circuit 132' is synchronously supplied with the above-mentioned three outputs (viz., the autocorrelation function R_hh'(|m_i - m_k|), the cross-correlation function φ_xh'(-m_k) and the selected grid q'), and determines an amplitude of each of the excitation pulses within one frame.
The decision circuit 502 is coupled to the circuits 132 and 132' to be supplied with the outputs: the autocorrelation functions R_hh(|m_i - m_k|) and R_hh'(|m_i - m_k|), the cross-correlation function φ_xh(-m_k) and φ_xh'(-m_k), and the selected grids q and q'. The decision circuit 502 determines power or energy J of an error signal between the incoming and reconstructed signals, according to the following equation (25), in connection with each of the two excitation pulse series which are obtained at the sections 508 and 506.
Equation (25) can be obtained by substituting equations (15) and (22) into equation (9). In equation (25), R_xx(0) represents power or energy of the output x_w(n) of the weighting circuit 136 (or 136').
Alternatively, the error signal energy can approximately be obtained using the following equation (26) instead of equation (25). ${J = Σ φ}^{2}_{xh} {(-m}_{i} {)/R}_{hh}$
Equation (26) utilizes an error of the cross-correlation function, which can be obtained by calculating the excitation pulse series.
The decision circuit 502 compares the two kinds of power or energy: one obtained depending on the parameters from the section 508 (referred to as Jo) and the other obtained depending on the parameters from the section 506 (referred to as Jo'). In the event of Jo'<Jo, the decision circuit 502 determines that the excitation pulse series obtained through the section 506 is suitable for use relative to that obtained through the section 508. In this case, the decision circuit 502 instructs the switch 500 to relay the output of the section 506 to the pulse coding circuit 137. Further, the decision circuit 502 opens the gate 504 allowing the coded information (coded-LAR(i), coded-Md and coded-β) to be applied to the multiplexor 300. In this case, the gate 504 attaches a predetermined code to the coded-Md and -β). Contrarily, in the event of Jo'>Jo, the decision circuit 502 forces the switch 500 to relay the output of the section 508 to the circuit 137, and opens the gate 504 to pass the above-mentioned coded information therethrough.
As shown, the two sections 506 and 508 are separately provided in the second embodiment. However, this invention is not limited to such an arrangement. That is to say, the impulse response calculating circuit 120 can be adapted to calculate the above-mentioned two functions h_w(n) and h_w'(n). In this case the circuit 120 generates hw'(n) by making zero the parameters Md' and β' which are applied to the coefficient weighting circuit 408. It goes without saying that h_w(n) is first calculated and thereafter computation of the h_w'(n) is performed or vice versa, which can be applied to the other blocks wherein two kinds of computation are implemented.
In the above-mentioned embodiments, various calculations can be carried out on a sub-frame basis as in the prior art.
The second embodiment can be modified such that the pitch gain β is compared with a predetermined threshold. If the pitch gain β is less than the threshold then the pitch gain β is rendered zero. This means that the excitation pulses are generated using the spectrum parameters only. It is understood that this modification no longer requires the provision of the decision circuit 502 and the calculations of equations (25) and (26). This variation can result in the reduced number of operations.
While the foregoing description describes only two embodiments of the present invention, the various alternatives and modifications possible without departing from the scope of the present invention, which is limited only by the appended claims, will be apparent to those skilled in the art.

Claims

An arrangement for encoding a speech signal using a regular pulse excitation scheme, comprising:
first means (112, 114, 116) for being supplied with a discrete-time speech signal and for dividing said discrete-time speech signal into a plurality of frames;

second means (118, 124) for extracting a plurality of parameters from each of said frames supplied by said first means;

synthesis means (122) for generating a signal using said plurality of parameters and a sequence of excitation pulses;

third means (120) for generating an impulse response function signal using said plurality of parameters;

fourth means (126) for generating an autocorrelation function signal using said impulse response signal; and

fifth means (128) for generating a cross-correlation function signal using said impulse response function signal and a weighted difference between one of said frames of discrete-time speech signal and one frame of said signal generated by said synthesis means;
characterized by:
sixth means (130) for generating a grid signal indicative of a location of a first excitation pulse within one frame using said cross-correlation function signal; and

seventh means (132) for receiving said autocorrelation function signal, said cross-correlation function signal and said grid signal, said seventh means determining an amplitude sequence of excitation pulses within one frame.
An arrangement as claimed in claim 1, wherein said second means (118, 124) comprises:
eighth means extracting one or more first parameters representative of a spectrum envelope from each of said frames supplied by said first means, encoding the first parameters, decoding the encoded first parameters and obtaining the decoded first parameters; and

ninth means extracting second and third parameters from each of said frames supplied by said first means, said second and third parameters being respectively representative of a pitch period and a pitch gain, said ninth means decoding the coded second and third parameters and obtaining the decoded second and third parameters,

wherein the decoded first, second and third parameters are applied to said third means (120).
An arrangement as claimed in claim 2, wherein said third means (120) comprises:
an impulse generator (400) for generating an impulse;

a long term prediction filter (402) receiving said impulse as well as said second and third parameters; and

a short term prediction filter (404) being coupled in series with said long term prediction filter (402) and receiving said first parameters and the output of said long term prediction filter.
A method for encoding a speech signal using a regular pulse excitation scheme, comprising the steps of:
(a) receiving a discrete-time speech signal and dividing said discrete-time speech signal into a plurality of frames;

(b) extracting a plurality of parameters from each of said frames of said discrete-time speech signal;

(c) generating a signal using said plurality of parameters and a sequence of excitation pulses;

(d) generating an impulse response function signal using said plurality of parameters;

(e) generating an autocorrelation function signal using said impulse response signal; and

(f) generating a cross-correlation function signal using said impulse response function signal and a weighted difference between one of said frames of discrete-time speech signal and one frame of said signal;
characterized by:
(g) generating a grid signal indicative of a location of a first excitation pulse within one frame using said cross-correlation function signal; and

(h) receiving said autocorrelation function signal, said cross-correlation function signal and said grid signal, and determining an amplitude sequence of excitation pulses within one frame.
A method as claimed in claim 4, wherein said step (b) comprises the steps of:
extracting one or more first parameters representative of a spectrum envelope from each of said frames of said discrete-time speech signal and encoding the first parameters, decoding the encoded first parameters and obtaining the decoded first parameters; and

extracting second and third parameters from each of said frames of said discrete-time speech signal wherein said second and third parameters are respectively representative of a pitch period and a pitch gain, and decoding the coded second and third parameters and obtaining the decoded second and third parameters,

wherein the decoded first, second and third parameters correspond to said plurality of parameters in said step (d).
A method as claimed in claim 5, wherein said step (d) comprises the steps of:
generating an impulse;

receiving said impulse as well as said second and third parameters, and generating an output representative of a pitch structure; and

receiving said first parameters and said output representative of a pitch structure, and generating an output representative of spectrum envelope characteristics.