Processing sequential patterns

(54) NAME Speech processor for generating voice message (56) PUBLICATIONS Cited: ---

<td>(57) SUMMARY:</td><td>A speech processor for providing voice messages. Here-</td>

at, a sequential sample processing arrangement operates to form an appropriately coded signal representing the sample. A sequential sample, for example a speech pattern, is divided into successive time intervals &lt;1191. In each interval, a group of signals is generated predictive of the interval sequential pattern, for example, a speech pattern and a signal, representing the difference between the interval quantial pattern (119) and the predictive signal (118). A first signal, corresponding to the interval pattern, is formed in response to said pattern predicative signals and said difference representative signal (121), and a second interval corresponding signal is generated in response to said pattern predicative signals (123), a signal corresponding to the differences between said first and second interval corresponding signals ( 125), is generated and a prescribed format encoded signal is provided in response to the interval difference corresponding signal (127), modifying the second signal to minimize the interval difference signal. Then the coded signal is utilized to construct a copy of the interval sequential pattern.

PRV 328 Al LF 130 9 132 AA

<img file="SE467429B_D0001.tif" />

The numbers mom parentheses enter^r internationei¹ identification koo INID-koc BoKStav mom Klammer indicates international document code

4o /

The present application relates to a speech processor for providing voice messages.

Digital voice communication facilities include audio storage and audio response aids that utilize signal compression to reduce the bit rate required for storage and / or transmission. As is well known in the art, a speech pattern contains redundancies, which are not essential to its good quality. Removing redundancy components from the speech pattern significantly reduces the number of digital codes required to construct a copy of the speech. However, the subjective quality of the talk copy depends on the compression and coding methods.

A well-known digital speech coding system, shown in U.S. Patent No. 3,624,302, includes linear prediction analysis of a speech input. The speech signal is divided into successive intervals and a group of parameters representing the interval number is generated. The parameter set includes linear prediction coefficient signals representing the spectral envelope of the speech in the range, as well as pitch and sound control signals corresponding to the speech excitation. These parameter signals can be encoded at a much lower bit rate than the speech signal waveform itself. A copy of the speech signal is formed by the parameter signal codes through synthesis. The synthesizing arrangement usually includes a model of the speech apparatus in which the excitation pulses are modified by the spectral envelope representative predictive coefficients of a predictive all-pole filter.

The previously pitch excited linear predictive coding is very effective. However, the talc copy produced exhibits a synthetic quality which is often difficult to understand. Generally, the low speech quality is obtained by the lack of correspondence between the speech pattern and the linear prediction model used. Errors in the pitch code or errors in determining whether a speech interval is pronounced tone or non-tone causes the speech copy to sound disturbed or unnatural. Similar problems also arise in format coding of speech.

467 429

Alternative coding arrangements, whereby the speech excitation is obtained from what remains after prediction, for example ADPCM or APC, provide a marked improvement, since the excitation does not depend on an inaccurate model. The excitation bit rate of these plants is at least one order of magnitude higher than that of the linear predictive model. Attempts to reduce the excitation bit rate in residual type plants have usually resulted in a significant deterioration in quality. It is an object of the invention to provide a speech processor which has better high quality speech coding at lower bit rates than at residual coding schemes.

We have found that the previous problems of residual coding can be solved by forming a pattern that is predictive of a pattern (e.g. speech pattern) to be encoded, and comparing the coding pattern with the predictive pattern on the frame prediction. frame basis. The differences between the coding pattern and the predicative pattern over each frame are used to form a coded signal of a prescribed format, which coded signal alters the predicative pattern to reduce the frame differences. The bit rate of the prescribed format encoded signal is selected in such a way that the modified predictive pattern approximates the speech pattern to a desired level which is compatible with the coding requirements.

The speech processor intended for solving the above problems includes the features specified in the characterizing part of the claim.

Z «7 I ö A

40 / 4Z&gt;

Drawing Description

Fig. 1 is a block diagram of a speech processor circuit illustrating the invention.

Fig. 2 shows a block diagram of an excitation signal-forming processor which can be used in the circuit of Fig. 1.

Fig. 3 shows a flow diagram illustrating the operation of the excitation signal-forming circuit of Fig. 1.

Figs. 4 and 5 show flow charts illustrating the operation of the circuit of Fig. 2.

Fig. 6 shows a timing diagram illustrating the operation of the excitation signal forming circuit of Figs. 1 and 2.

Fig. 7 shows waveforms illustrating the speech processing according to the invention.

Detailed description

Fig. 1 shows a general block diagram of a speech processor illustrating the invention. In figure 1, a speech pattern is received, e.g. a spoken message, by the microphone converter 101. The corresponding analog speech signal therefrom is band-limited and converted into a sequence of pulse samples in the filter and sampler circuit 113 of the prediction analyzer 110. The filtration can be arranged to remove frequency components of the speech signal above 4.0 kHz and the sampling can take place at a frequency rate of 8.0 kHz, as is well known in the art. The timing of the samples is controlled by the sample clock CL from the clock generator 103. Each sample from the circuit 113 is transformed into an amplitude representative digital code in the analogue / digital converter 115.

The sequence of speech samples is fed to the predictive parameter computer 119 which, as is well known in the art, works to divide the speech signals into 10-20 ms intervals and to generate a group of linear prediction coefficient signals aj, k = 1.2, ..., p which represents the predicted short time spectrum of the N &gt;&gt; p number samples in each interval. The speech samples from the analogue / digital converter 115 are delayed in the delay device 117 to allow time for the generation of signals aj. The delayed samples are fed to the input of the prediction residual generator 118. The predicted residual generator responds, as is well known in the art, to the delayed speech samples and prediction parameters a ^ to form a signal corresponding to

467 429 the difference between them. The formation of the predictive parameters and the prediction residual signal for each frame shown in the predictive analyzer 110 can be performed according to the arrangement disclosed in U.S. Patent 3,740,476 or in other well-known arrangements in the art.

Although the predictive parameter signals a ^ form an appropriate representation of the short-term speech spectrum, the residual signal usually varies greatly from range to range and exhibits a high bit rate, which is unsuitable for many applications. In the pitch excited vocoder, only the peaks of the remainder are transmitted as pitch pulse codes. However, the resulting quality is usually poor. The waveform 701 in Figure 7 illustrates a normal speech pattern over two time frames. The waveform 703 shows the predictive residual signal obtained from the waveform 701 pattern and the predictive parameters of the frames. As is readily apparent, waveform 703 is relatively complex so that coding of pitch pulses, corresponding to the peaks therein, does not provide adequate approximation. mingling of the predictive remnant. According to the invention, the excitation code processor 120 receives the residual signal d d and the prediction parameters a ^ of the frame and generates an interval excitation code having a predetermined number of bit positions. The resulting excitation code, shown in waveform 705, exhibits a relatively low bit rate which is constant. A copy of the speech pattern in waveform 701, constructed from the excitation code and the prediction parameters of the frames, is shown in waveform 707. As can be seen in a comparison of waveforms 701 and 707, more high-quality speech characteristics of adaptive predictive coding are obtained at much lower bit rates.

The prediction residual signal d 1 and the predictive parameter signals a 2 for each subsequent frame are transmitted from circuit 110 to excitation signal forming circuit 120 at the beginning of the subsequent frame. Circuit 120 operates to produce a multi-element frame excitation code EC having a predetermined number of bit positions for each frame. Each excitation code corresponds to a sequence of 1 &lt;i &lt;1 pulses representing the excitation function of the frame. The amplitude and position m 2 of each pulse in the frame are determined in the excitation signal forming circuit to allow the construction of a copy of the frame speech signal from the excitation signal and the frame predictive parameter signals. (3_χ- and πι_χThe signals are encoded in encoder 131 and multiplied by the frame's

467 429 prediction parameter signals in the multiplexer 135 to provide a digital signal corresponding to the frame number pattern.

In the excitation signal forming circuit 120, the predictive residual signal d 2 and the predictive parameter signals a 2 of a frame are applied to the filter 121 via the gates 122 and 124. At the beginning of each frame, the frame clock signal FC opens gates 122 and 124, whereby the d signals to filters 121 and 123. Filter 121 is adapted to modify signal d 2 so that the quantization spectrum of the error signal is concentrated in its formant ranges. As shown in U.S. Patent 4,133,976, this filter arrangement is suitable for masking the error in the high signal energy lines of the spectrum.

The transfer function of filter 121 is expressed in z-transformation form as ^HCz-^{) =} 1B (z) where B (z) is controlled by the frame predictive parameters * a ^.

The predictive filter 123 receives the frame predictive parameter signals from the computer 119 and an artificial excitation signal EC from the excitation signal processor 127. The filter 123 has the transfer function of Equation 1. The filter 121 forms a weighted frame number signal y which reacts to the predictive residue d , which responds to the excitation signal from the signal processor 127. The signals y and y are correlated in the correlation processor 125, which generates a signal E corresponding to the weighted difference therebetween. The signal E is sent to the signal processor 127 for adjusting the excitation signal EC, so that the differences between the weighted speech representative signal from filter 121 and the weighted signal from filter 123 representing artificial speech are reduced.

The excitation signal is a sequence of 1 &lt;i &lt;_ I pulses. Each pulse has an amplitude and a position n The processor 127 is intended to successively generate&gt; m 2 signals which reduce the difference between the weighted speech representative frame signal from filter 121, and the weighted frame signal from filter 123, which represents artificial speech. The weighted speech representative frame signal can be expressed as:

y_n = ? ^dk^hrsnk ¹ &lt; ⁿ &lt; ^N⁷ⁿ k = nk ^{K n K} (2)

467 429 and the weighted signal of the frame, representing artificial speech, can be expressed as:

** iλ i ^ n = E j nm. 1 &lt;η &lt;N (3) j = 1 where h_R is the pulse response of filter 121 or filter 123.

The excitation signal formed in circuit 120 is an encoded signal with the elements (L, m., I = 1.2, ..., 1. Each element represents a pulse in the time frame. 0 ^ is the amplitude of the pulse and now is the position of the pulse. The correlation signal generator circuit 125 acts to successively generate a correlation signal for each element, each element being located at time 1 &lt;q &lt;Q in the time frame. Accordingly, the correlation processor circuit Q forms possible candidates for elements in accordance with Equation 4

<img file="SE467429B_D0002.tif" />

N

C n = q

<img file="SE467429B_D0003.tif" />

<img file="SE467429B_D0004.tif" />

<img file="SE467429B_D0005.tif" />

Σ n = q

7_n, ii nq (4) where y.

⁷n, i-1 i-1

E. j = 1

ph ^{3 n}‘^mj (5)

127 receiver (Lq signals)

The excitation signal generator from the correlation signal generator circuit and selects it

C- signal, which has the maximum absolute value, and forms it

A in the tenth element of the coded s, the ignal

<img file="SE467429B_D0006.tif" />

ιη_± = A * (6) where q * is the position of the correlation signal which has the maximum absolute value. The index i is the addition to i + 1 and the signal y_n at the output of the predictive filter 123 is modified. The process of Equations 4, 5 and 6 is repeated to form the elements^mIn the formation of the elements m 2, the signal having the elements m 2 is transmitted to the encoder 131. As is well known in the art, encoder 131 works to quantize (the Lnu elements and to form an encoded signal suitable for transmission to the network 140.

467 429

Each of the filters 121 and 123 of Fig. 1 may comprise a transversal filter of the type described in the aforementioned U.S. Patent No. 4,133,976. Each processor 125 and 127 may comprise one of the processor arrangements well known in the art and intended to perform the processing required by Equations 4 and 6, e.g. CSP, Inc. Macro Arithmetic Processor System 100 or other well-known processor arrangement.

Processor 125 includes a read-only memory that permanently stores programmed instructions for controlling the C. signal formation J · H.

in accordance with Equation 4, and the processor 127 includes a read-only memory which permanently stores programmed instructions for selecting the m 2 signal elements of Equation 6, as is well known in the art. The program instructions in processor 125 are given in FORTRAN language form in Appendix A and the program instructions in processor 127 are given in FORTRAN language form in Appendix B.

Fig. 3 shows a flow chart illustrating the operation of processors 125 and 127 for each time frame. Referring to Fig. 3, the high-impulse response signals in box 305 are generated in response to the frame predictive parameters of the transfer function in Equation 1. This occurs upon receipt of the FC signal from clock 103 in Fig. 1, according to left pane 303. The element index i and excitation pulse position index £ are initially set to 1 in box 307.When receiving the signals y_n and y_n from the predictive filters 121 and 123, the signal is generated according to box

309th The position index £ is added in box 311 and the formation of the next position signal is initiated.

After the C signal is formed for the excitation signal element i in processor 125, processor 127 is activated.

The £ index in processor 127 is initially set to 1 in box 315 and in. The index and the C signals formed in processor 125 are transmitted to processor 127. The CL CL * signal, which represents the C den signal, which has the maximum absolute value, and its position q * is set to zero in box 317. The absolute values ​​of the

<img file="SE467429B_D0007.tif" />

the ignals are compared so that absolute values ​​are stored which include the boxes 319, with the signal C

321 , 323 and 325 µm maximum of in that loop,

After the C signal from processor 125 has been processed, transfer from box 325 to box 327 takes place.

467 The 429 code element position is set to q * and the excitation element size is generated according to Equation 6. (The m 2 element is output to the predictive filter 123 according to box 329 and the index i is added according to box 329. When forming the frame element, new transfer from the waiting box 303 is made from the decision box. 331. The processors 125 and 127 are then placed in a wait state until the next frame's FC frame clock pulse.

The excitation code in the processor 127 is also fed to the encoder 131. The encoder works to transform the excitation code from the processor 127 into a form suitable for use in the network 140. The prediction parameter signals a ^ of the frame are transmitted to an input of the multiplexer 135 via the delay device 133 as prediction signals. The excitation coded signal ECS from encoder 131 is transmitted to the second input of the multiplexer. The multiplied excitation and predictive parameter codes from the frame are then transmitted to the network 140.

The network 140 may be a communication facility, the message memory of an audio storage arrangement, or an apparatus intended to store a complete message or a vocabulary of prescribed message units, e.g. words, phonemes, etc for use in speech synthesizers. Whatever the message unit, the resulting sequence of frame codes is advanced from the circuit 120 via the network 140 to the speech synthesizer 150. The synthesizer utilizes, in turn, the frame excitation codes from circuit 120 as well as frame predictive parameter codes to construct a copy of the speech pattern.

Demultiplexer 152 in synthesizer 150 separates excitation code EC in a first row from its prediction parameters a ^ · After the excitation code has been decoded in an excitation pulse sequence in decoder 153, the code is sent to the excitation input of speech synthesizer filter 1544. the a ^ codes are transmitted to filter 154. The filter 154 operates in response to the excitation and predictive parameter signals to form an encoded copy of the frame number signal, as is well known in the art. The digital / analog converter 156 is intended to transform the encoded copy into an analog signal which is passed through the low pass filter 158 and transformed into a speech pattern via the converter 160.

An alternative arrangement for performing the excitation code formation operations of circuit 120 may be based on the weighted mean square error between the signals square error at the forming signal pulse is off (T and

Λ ί. Ί / OQQ u / t c-&gt; y_n and y. This weighted mean for the i nth excitation N

E = S ¹ n = 1

ph

J n-rnj (7) is the nth jth pulse where hn

for that of the jth pulse.

The pulse modes and Excitation 7th equation 7 to a sample of in the excitation code signal and the pulse response of H (z), h1 is the position, the strength of the pulse amplitudes is sequentially generated, elements are determined by the minimum. Equation 7, reduction can be rewritten by E. ii as

<td>K E. = Σ¹ n = 1</td><td>Pn- Ä \ j = 1</td><td>13, h 3 pm</td><td colspan="2"> \² + / 3? H² &lt;· j) x nm.</td>

<td></td><td>-2/3. in:</td><td>fh</td><td>i-1, - Σ β .hh \</td><td> (9)</td>

<td></td><td>'in V</td><td>n nm. 1</td><td> 3 = 1 ^{3 n}^mj ^ i)</td><td></td>

so that the known excitation code elements preceding (3 ^) appear only in the first expression.

As is well known, the value of decreasing E E to a minimum can be determined by differentiating Equation 8 with respect to (3 ^ and by setting dB,

<img file="SE467429B_D0008.tif" />

(9)

Consequently, the optimal value is

<img file="SE467429B_D0009.tif" />

tn £ + K

Σ k = nL · -K i-1

<img file="SE467429B_D0010.tif" />

<img file="SE467429B_D0011.tif" />

(10)

<img file="SE467429B_D0012.tif" />

<img file="SE467429B_D0013.tif" />

(11)

<img file="SE467429B_D0014.tif" />

467 429, the autocorrelation coefficients of the predictive filter pulse response signal h

(3 ^ in Equation 10 is a function of the pulse site and is determined for each possible value thereof. The maximum of the values ​​over the possible pulse sites is then selected. After obtaining the values ​​of β. And m., The values ​​β. And m. .

1 + 1 1 + 1 by solving Equation 10 in the same way. The first expression in Equation 10, ie^mk^{+ K}k = m? -K corresponds to the speech-representative signal of the frame at the output of the predictive filter 121. The second term in equation 10, ie i-1 £ Μ j = 1 r^mj^mi 'corresponds to the signal of the frame representing artificial speech, at the output of the predictive filter 123, the amplitude of an excitation pulse at the position m 2, which minimizes the difference between the first and second terms.

The data processing circuit shown in Fig. 2 provides an alternative arrangement for the excitation signal forming circuit 120 of Fig. 1. The circuit of Fig. 2 provides the excitation code for each frame of the speech pattern in response to the frame prediction residual signal d ^ and the frame prediction parameter signals a ^ in accordance with equation 10 and may comprise it. The prior art arrangement of CSP, Inc. Macro Arithmetic Processor System 100 or other processor arrangements, as is well known in the art.

As shown in Fig. 2, processor 210 receives the predictive parameter signals a, and the prediction residual signals d of each subsequent frame of the speech pattern from the circuit 110 via the memory 218. The processor operates to form the excitation code signal elements nip. ^m£’ ’**’ ^1’ ^mIN ^{un &lt;}^^your control of permanently stored instructions in predictive filter subprogram memory 201 and excitation processing subprogram memory 205. The predictive filter subprogram of memory ROM 201 is set forth in Appendix C and excitation processing subprogram of memory ROM 205 is set forth in Appendix D.

Processor 210 includes common bus line 225, data memory 230, central processor 240, arithmetic processor 250, control interface 220, and input-output interface260.

467 429

As is well known in the art, central processor 240 is intended to control the frequency of operations of other units of processor 210 in response to coded instructions from controller 215. Arithmetic processor 250 is intended to control the arithmetic processing of coded signals from data memory 230 in response to control signals from the central processor 240. The data memory 230 controls signals routed through the central processor 240 and provides these signals for the arithmetic processor 250 and the input / output interface 260. The control interface 220 provides a communication link for the program instructions in the memory ROM 201 and the memory ROM 205 to the central processor 240 through the controller 215. the input-output interface 260 allows the signals d 1 and a 3 to be fed to the data memory 230 as well as to output outputs β. and m- from the data memory to the encoder 131 in Fig. 1.

The operation of the circuit of FIG. 2 is illustrated in the filter parameter processing flow diagram of FIG. 4, the excitation code processing flow diagram of FIG. 5, and the timing diagram of FIG. r is set for the first frame by a single pulse ST from clock generator 103. Fig. 6 illustrates the operation of the circuit of Figs. 1 and 2 for two consecutive frames. Between the times t_Q and because in the first frame, the prediction analyzer 110 forms the speech pattern samples of frame r + 2 as in waveform 605 under the control of sample clock pulses with waveform 601. Analyzer 110 generates the α + signals corresponding to frame r + 1 between the times t_Q and tj and form the predictive residual signal d between the times t ^ and t ^, as indicated in waveform 607. The signal FC (waveform 603) occurs between the times ΐθ and t ^. The signals d 1 from the residual signal generator 118 previously stored in memory 218 during the previous frame are placed in the data memory 230 via the input-output interface 260 and the common bus line 225 under the control of the central processor 240. As indicated by the working box 415 in FIG. 4 these operations respond to the frame clock signal RC. The frame prediction parameter signals a 1 from the prediction parameter computer 119, which were previously placed in memory 218 during the previous frame, are also inserted into memory 230 according to work box 420. These operations occur between the times

<img file="SE467429B_D0015.tif" />

<img file="SE467429B_D0016.tif" />

467 429 and t in Figure 6.

After inserting the frame signals and α 1 into the memory

230 one goes to box 425 and the predictive filter coefficients b ^, which correspond to the transfer function in equation 1, ^Akk - 1.2, ..., p (12) are generated in arithmetic processor 250 and placed in data memory 230. £ is normally 16 and tx is normally 0.85 for a sampling rate of 8 kHz. Predictive filter impulse response signals

<img file="SE467429B_D0017.tif" />

<img file="SE467429B_D0018.tif" />

<img file="SE467429B_D0019.tif" />

min (k-1, p)

<img file="SE467429B_D0020.tif" />

is then generated in arithmetic processor 250 and stored in data memory 230. When the h 2 pulse response signal. is stored, you go to box 435 and the predictive filter autocorrelation signals of equation 11 are generated and stored.

At time t₂ in Fig. 6, controller 215 switches memory ROM 201 from interface 220 and connects excitation processing subprogram memory ROM 205 to interface. The formation of the excitation pulse codes and now, shown in the flow diagram of Fig. 5, is then initiated. Between the times t₂ and t₄ in Figure 6, the excitation pulse sequence is formed. The excitation pulse index i is initially set to 1 and the pulse position index £ is set to 1 in box 505. is set to zero in box 510 and one goes to operation box 515 to determine is the optimal excitation pulse at position q = 1 of the frame. The absolute value of 0 is then compared with the previously stored value of in decision box 520. Since initially it is zero, the m ^ code is set to q = 1 and the β ^ code to in box 525.

The position index £ is then added in box 530 and one goes to box 515 via decision box 535 to generate the signal β₂The loop comprising the boxes 515, 520, 525, 530 and 535 is iterated for all heart rate values ​​1 &lt;q &lt;Q. After the Qth iteration, the first excitation pulse amplitude ^ 1 is stored ⁼ @ Iq * ^och ^^{ace read} in ^{ramen m}in ⁼ q * i ^min this way, the first of the excitation pulses is determined. During

467 429 reference to waveform 705 in Fig. 7, the frame r occurs between the times ΐθ and t och. The excitation code for the frame consists of 8 pulses. The first pulse with the amplitude f 1 and the position m 2 occurs at the time t in Fig. 7, as determined in the flow chart of Fig. 5 for index i = 1.

Index i is added to the subsequent excitation pulse in box 545 and continues to operation box 5.15. via box 550 and box 510. At the end of each iteration of the loop between boxes 510 and 550, the excitation signal signal is modified to further reduce the signal in Equation 7. At the end of the second iteration, the pulse is generated.₂ (time t_m2 in waveform 705). The excitation pulses µm (time t_m3) , %^m4 (time, fym ^ (time t 5), p ^ m ^ (time tm6 ^&gt; fy^m7 (time t_my) and fymg (time t is then successively formed as index i. is added).

After the Ith iteration (waveform 609 at t₄) continue to box 555 from decision box 550 and the current frame excitation code fym.], ^₂^m2^} - »fy^mJG^eneseras therein. The frame index is added in box 560 and the predictive filter operations of Figure 4 for the next frame are started in box 415 at the time ty in Figure 6. Upon occurrence of the FC clock signal for the next frame at ty in Figure 6, the predictive parameter signals for the frame r + 1 are formed (waveform 605 between the times ty and t ^) and a ^, and the d ^ signals are generated for the frame r + 2 (waveform 607 between the times ty and fi 3) °^t12'⁾ *

The frame excitation code of the processor of Figure 2 is supplied via the input-output interface 260 to the encoder 131 of Figure 1, as is well known in the art. The encoder 131 operates as previously mentioned to quantize and format the excitation code for application to the network 140. The α-prediction parameter signals of the frame are transmitted to an input of the multiplexer 135 via the delay device 133, so that the frame excitation code of the encoder 131 can be appropriately multiplied therewith.

The invention has been described with reference to particular illustrative embodiments. It will be apparent to those skilled in the art that various modifications may be made without departing from the scope of the invention. For example, the embodiments described herein have utilized linear predictive parameters as well

467 429 a predictive residue. The linear predictive parameters can be replaced by formant parameters or other number parameters well known in the art. The predictive filters are then arranged to respond to the speech parameters utilized and to the speech signal so that the excitation signal formed in circuit 120 of Fig. 1 is used in combination with the speech parameter signals to construct a copy of the speech pattern of the frame in accordance with the invention. The coding arrangement according to the invention can be extended to sequential patterns, e.g. biological and geological patterns, to obtain appropriate representations thereof.

467 429