JPH0738117B2 - Multi-pulse encoder - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a multi-pulse coding device, and more particularly to a speech signal to be reproduced in a multi-pulse coding device including a pitch prediction loop to determine an optimum multi-pulse train. The present invention relates to a multi-pulse encoding device that significantly reduces the influence of quantization noise on the.

[Conventional technology]

It is well known that a multi-pulse coding apparatus that includes a pitch prediction loop and configures the analysis side to improve the ratio S / N between the input speech power S and the reproduced speech quantization noise N by expecting the pitch prediction gain. is there.

Pitch prediction methods in a conventional multi-pulse coding apparatus using this type of pitch prediction loop are roughly classified into the following three types.

(1) A pitch prediction method in which a pitch prediction system and a multi-pulse processing system are handled independently of each other.

(2) A pitch prediction method in which a proximity prediction system and a pitch prediction system for spectrum envelope parameter analysis are configured in a cascade connection form.

(3) A pitch prediction method in which a multi-pulse coding system is inserted in the pitch prediction loop.

Here, the principle of pitch prediction will be described first.

FIG. 7 is a block diagram showing a general structure of the pitch predictor.

The pitch predictor is configured as a kind of digital filter, and basically provides filter coefficients that should obtain an output as similar as possible to the input, β ₁ to β ₃ in the case of FIG. 7, to the digital filter each time a voice input is received. While forward (forw
ard) to obtain a voice output as an optimum prediction signal. The method of calculating the pitch prediction coefficient itself will be described in detail later.

The input is either a voice signal or a sound source as a vocal cord vibration waveform. Also, as a whole, as shown in FIG.
It has also been clarified from many experimental results that it is possible to secure an optimum predicted signal which is practically almost satisfactory under the condition of determining a filter including two points separated by a period.

The variable delay circuit 55 gives a delay of T _P −1 period to the input based on the pitch period T _P obtained from the previous frame analysis information, and as a result, the unit delay elements 56-1 and 56-2 are provided.
The output point time information is retention time status of each T _P and T _P +1 week period. The outputs of these three points are added by the adders 58 after being multiplied by β _{1 to} β ₃ by the multipliers 57-1 to 57-3, respectively, and output. It is a prediction output in the prediction format. Next, three types of the above-described pitch prediction methods performed by using such a pitch predictor will be described.

FIG. 4 is a block diagram showing a multi-pulse code decoding system in which the pitch prediction system and the multi-pulse processing system are independent of each other.

The output of the pitch predictor 34 on the analysis side is subtracted from the speech input by the adder 35 and supplied to the multi-pulse encoder 36 as the pitch prediction residual.

The multi-pulse encoder 36 searches for and encodes multi-pulses based on the input pitch prediction residual, and supplies them to the multi-pulse decoder 37 on the analysis side. The multi-pulse decoder 37 decodes the input to restore the multi-pulse train, and then supplies it to the adder 38. The adder 38 has a pitch predictor 39
Is also supplied, and a speech output is obtained in a format in which the predicted waveform generated by the pitch predictor 39 is successively added and output based on the output of the adder 38.

FIG. 5 is a block diagram showing a multi-pulse code decoding system in which the pitch prediction system is cascade-connected to the proximity prediction system. Pitch predictors 41 and 45 are respectively arranged on the analysis side and the synthesis side in a form of cascade connection to the proximity prediction filters 42 and 46. The proximity prediction filters 42 and 46 are used for LPC (Linear Prediction Codin) between several samples that are sufficiently close to each other compared to pitch search.
g, linear predictive coding) coefficient is extracted for each predetermined analysis frame for the predicted waveform provided from the pitch predictors 41 and 45. The adder 43 takes the proximity prediction error between the output of the proximity prediction filter 42 and the voice input, and supplies this to the error analyzer 44. The error analyzer 44 provides input to the multi-pulse encoder 40 while analyzing the input spectral and power levels, which multi-pulse encoder 40 provides white noise and power to the spectrum of the proximity prediction error waveform thus provided. Is obtained and provided as pitch source information to the pitch predictors 41 and 45, and a speech output is obtained from the proximity prediction filter 46.

FIG. 6 is a block diagram showing a multi-pulse code decoding system in which a multi-pulse coding system is inserted in the pitch prediction loop. The pitch prediction method in this case includes a multi-pulse coding processing part in the pitch prediction loop on the analysis side, and the influence of the quantization noise generated via the encoder on the reproduced voice is negative feedback of the entire analysis system. It is hoped that it was alleviated by the action of the loop.

In FIG. 6, the voice input is x, and the adder 47 is a pitch predictor.
It is represented by the predicted waveform y supplied from 51.

The input of the multi-pulse encoder 48 is x-y, and the multi-pulse encoder 48 and the multi-pulse decoder 49 perform multi-pulse search coding and decoding on x-y to obtain the multi-pulse. The output − is output from the decoder 49. The reason for the existence of the multi-pulse decoder 49 included in the analysis side is to equalize the conditions for the synthesis side and the quantization noise.

Now, the-is fed to the adder 50, which results in the pitch estimator 51 being provided with a + y- input. The above-is xy which includes the quantization noise in the encoding and decoding, and the pitch predictor 51 obtains the predicted waveform y based on approximately since y- is sufficiently smaller than Pitch prediction is performed and this is supplied to the adders 47 and 50.

xy is obviously the pitch prediction residual, and the pitch prediction residual is improved in its S / N ratio due to the effect of the negative feedback loop for the multi-pulse coding / decoding system.

On the combining side, the multi-pulse decoder 52 restores −,
This and y obtained from the pitch predictor 54 are added by the adder 53,
Outputs + y-. This is approximately equal to and the improved S / N with the negative feedback loop.
Is an audio input x containing the quantization noise of.

As a conventional pitch prediction method, the effect of the quantization noise in the encoder and decoder is theoretically mitigated by the effect of the negative feedback (hereinafter simply referred to as "feedback") shown in FIG. It has become the mainstream with the expectation that.

[Problems to be solved by the invention]

In the multi-pulse coding apparatus including the conventional pitch prediction loop described above, the multi-pulse is determined independently of the feedback loop in the process of determining the multi-pulse. This is true for many APCs, including pitch prediction loops, for example.
As can be seen in (Adaptive Predictive Ccding) method, etc., when performing multi-pulse search, block processing is performed in units of tens of msec of analysis frames while extracting input speech every tens of msec, while pitch prediction is A pitch that should be one ahead of one pitch is predicted in a so-called forward manner. Therefore, there is a drawback in that the effect of improving the quantization noise of the pitch prediction feedback loop for multi-pulses is practically eliminated.

An object of the present invention is to eliminate the above-mentioned drawbacks and to provide a multi-pulse encoding device which can surely guarantee the influence of the S / N improvement effect by the feedback loop of the pitch prediction system.

[Means for solving problems]

The apparatus of the present invention is, in a multi-pulse encoding apparatus having a pitch prediction loop, preliminarily preliminary tentatively deciding a multi-pulse train consisting of settings larger than a necessary number, and then performing pitch prediction on the tentatively decided multi-pulse train The optimum multi-pulse train consisting of the required number of pulses is determined along with the pulse amplitude quantization level.

〔Example〕

 The present invention will now be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the analysis side of the multi-pulse coding apparatus of the present invention.

The configuration of the embodiment shown in FIG. 1 has an H '(Z) / H (Z) converter 1, an LPC analyzer 2, a K quantizer / decoder 3, a K / α converter 4, and an attenuation coefficient calculator. 5, pitch predictor 6, pitch prediction coefficient calculator 7, memory 8, adder 9, memory 10, quantizer / decoder 11, multi-pulse preliminary analyzer 12, pulse amplitude quantizer 13, memory 14, DP (Dynamic Programming) Processor 5, adder 16, LPC synthesis filter 1
7, an adder 18, an error power calculator 19, an adder 20, a pitch predictor 21, a multiplicator 22 and the like.

The H ′ (Z) / H (Z) converter 1 is a so-called perceptual weighting filter, and is a filter of a transfer function H (Z) having a filter coefficient of the α parameter received from the K / α converter 4.
It is converted into a filter, and a voice input is passed through the filter to perform noise shaping as preprocessing. Here, γ is an attenuation coefficient, which is set in the range of 0 <γ <1 and is provided from the attenuation coefficient calculator 5 in the form of γα. In this noise shaping, the quantization noise is fed back to process the spectrum of the quantization noise so as to be close to the spectrum of the voice signal, and the auditory noise is reduced by the masking effect. In other words, the characteristics of the filter formed by γα are similar to the spectral envelope of the voice input, and the quantization noise spectrum of the input voice signal is approximated to the spectrum envelope of the voice signal to achieve colorization, which depends on the human auditory characteristics. It is intended to effectively reduce noise.

The transfer function of the perceptual weighting filter by the H '(Z) / H (Z) converter 1 is represented by W (Z) in the following equation (1).

In equation (1), P is the analysis order.

Now, the LPC analyzer 2 performs LPC analysis on the speech input for each predetermined analysis frame, outputs a P-order K parameter as an LPC coefficient, and supplies this to the K quantizer / decoder 3.

The K quantizer / decoder 3 quantizes the input K parameter, decodes it again, and supplies it to the K / α converter 4. K
Such processing in the quantizer / decoder 3 is performed for the purpose of setting the condition of the quantization error generation on the analysis side to be substantially the same as that on the synthesis side.

The K / α converter 4 converts the P-th order K parameter into the K-th order α parameter, and the attenuation coefficient calculator 5 and H ′ (Z) / H.
(Z) Provide to the converter 1.

The γα output from the attenuation coefficient calculator 5 is also supplied to the multi-pulse preliminary analyzer 12 and the LPC synthesis filter 17.

γα is obtained by multiplying the P-order α parameter obtained for each analysis frame of the LPC coefficient of the voice input by the attenuation coefficient γ (0 <γ <1) in the perceptual sound weighting process.
It is the same as the LPC coefficient. Multi-pulse preliminary analyzer 12
Is a known processing method using the LPC coefficient received in the form of γα, and in this embodiment, the correlation region processing is used to search the multi-pulse train for each analysis frame.

FIG. 8 is a basic block diagram of the multi-pulse search by the correlation region processing, and the multi-pulse preliminary analyzer 12 also searches the multi-pulse train with the processing contents of FIG.

Now, letting ε be the difference between the synthesized signal synthesized by K multi-pulses and the voice input, ε is expressed by the following equation (2).

In the equation (2), N is the analysis frame length, and gi and mi are the amplitude and position of the i-th pulse in the analysis frame, respectively. The pulse amplitude and position that minimize ε are (2)
It is determined as maximizing the following expression (3) obtained by partially differentiating the expression with respect to gi and setting it to zero.

In the equation (3), Rhh is an autocorrelation coefficient of the impulse response of the voice synthesis filter, and hs is a cross-correlation coefficient of the voice input waveform and the impulse response. The content of the equation (3) is that gi (mi) is optimum as the amplitude when a pulse is generated at the time position mi. Also, gi
To obtain (mi), the numerator on the right side of equation (3) is calculated each time a pulse as a multi-pulse is determined.
The value of hs (mi) is corrected and then normalized by the autocorrelation coefficient Rhh (0) at the delay time of zero, and the maximum absolute value is searched for next. In this process, the second term of the numerator of the equation (3) as the correction value of the cross-correlation coefficient is the amplitude gl of the maximum value and the time position information me retrieved immediately before, and the delay time from the maximum value | me-mi. It is obtained based on the autocorrelation coefficient Rhh (| me-mi |) at |, the position information in the analysis frame of the multipulse to be searched, and the like.

The impulse response calculator 59 in FIG. 8 inputs γα to obtain an impulse response, and the autocorrelation coefficient calculator 60 calculates the autocorrelation coefficient. Further, the cross-correlation coefficient calculator 61 calculates the cross-correlation coefficient between the input waveform and the impulse response.

The multi-pulse searcher 62 uses these data (3)
The multi-pulse train for each analysis frame is provisionally determined based on the formula. Note that the number of multi-pulses provisionally determined in this way is set to be larger than the necessary number for the DP processing described later.

Returning to FIG. 1 again, the description of the embodiment will be continued. The multi-pulse pre-analyzer 12 retrieves a multi-pulse train for each analysis frame by using γα as the LPC coefficient and the pitch prediction residual waveform output from the adder 9 as the input waveform, and pulse amplitude quantization Supply to the vessel 13.

The pitch predictor 6 inputs the speech signal after the weighting of the hearing sound output from the H ′ (Z) / H (Z) converter 1 and the decoded pitch prediction coefficient provided from the quantizer / decoder 11. While generating the predicted waveform, the predicted waveform is output to the adder 9. As a result, the pitch prediction residual waveform is supplied from the adder 9 to the multi-pulse preliminary analyzer 12 and also stored in the memory 10.

The pitch prediction coefficient calculator 7 calculates a pitch prediction coefficient necessary for pitch prediction. As described with reference to FIG. 7, _three pitch prediction coefficients are usually determined by β ₁ to β ₃ which are filter coefficients of the pitch prediction filter.

The pitch prediction coefficient is calculated on the basis of the following calculation roots.

Now, assuming that the integer closest to the pitch period is Q, the following expression (4) is established.

e _{i + Q} + S _{i + Q} = β ₁ S _{i + 1} + β ₂ S _i + β ₃ S _i-1 (4) In formula (4), e _{i + Q} is the pitch prediction residual at the position of sample time i + Q. Difference waveform, S _{i + Q} is a pitch prediction value, waveforms based on S _{i + 1 to} S _i-1 are original speech waveforms at positions i + 1, i, i−1, β ₁ + β ₃ are pitches with respect to these original speech waveforms It is a prediction coefficient. That is, the equation (4) shows the relationship between the predicted waveform and the predicted value for the original waveform.

Β between the original waveform and the waveform to be predicted based on equation (4)
Considering the correspondence of 3 points via ₁ ~ β ₃ , the following (4) ~
Formula (6) is materialized.

e _{i + Q}・ S _{i + 1} ＋ S _{i + Q}・ S _i ＝ β ₁ S _{i + 1}・ S _{i + 1} ＋ β ₂ S _i・ S _{i + 1} ＋
β ₃ S _i-1・ S _{i + 1} (4) e _{i + Q}・ S _i ＋ S _{i + Q}・ S _i ＝ β ₁ S _{i + 1}・ S _i ＋ β ₂ S _i・ S _i ＋ β ₃ S
_i-1・ S _i …… (5) e _{i + Q}・ S _i-1 ＋ S _{i + Q}・ S _i-1 ＝ β ₁ S _{i + 1}・ S _i-1 ＋ β ₂ S _i-1・ S
_i-1 + β ₃ S _i-1 · S _i-1 (6) It is assumed that the original speech waveform has stationarity, and that the pitch prediction residual and the original speech waveform are uncorrelated. .
This assumption is an assumption that has practically no problem in speech processing.

By the way, the equations (4), (5) and (6) show the relational expressions between the original speech waveform and the waveform to be reproduced via the _three pitch prediction coefficients β ₁ , β ₂ and β _3. However, the two waveforms are related by an equation based on the waveform multiplication value at the corresponding time between the two waveforms. Unknown β _{1 to be obtained} ,
β ₂ and β ₃ must be such that the power error between the original speech waveform represented by these three equations and the reproduced waveform is minimized. This solution is found by finding a combination of β ₁ to β ₃ that makes the difference between the right side and the left side of the equations (4), (5), and (6) zero, and by applying the method of least squares. Easy to find. However, since (4), (5), and (6) are expressed in the form of vector products of waveform multiplication, they are once converted into expression by voice power and the least square method can be applied.

Waveform multiplication is similar to taking autocorrelation in this case. Therefore, equations (4), (5) and (6) are integrated for i and converted into the following equations (7), (8) and (9). sell.

ρ _Q-1 = β ₁ ρ ₀ + β ₂ ρ ₁ + β ₃ ρ ₂ …… (7) ρ _Q = β ₁ ρ ₁ + β ₂ ρ ₀ + β ₃ ρ ₁ …… (8) ρ _{Q + 1} = β ₁ ρ ₂ + β ₂ ρ ₁ + β ₃ ρ ₀ (9) In equations (7), (8) and (9), ρ _Q-1 , ρQ,
ρ _{Q + 1} is the autocorrelation coefficient of the waveform to be reproduced on Q-1, Q and Q + 1, respectively, and ρ ₀ , ρ ₁ , and ρ ₂ are i + 1,
It is an autocorrelation coefficient of the original speech waveform at i, i-1.

The following equation (10) is obtained from equations (7), (8) and (9), and β ₁ to β ₃ can be easily obtained.

The pitch prediction coefficient analyzer 7 uses the input weighted speech signal, calculates the pitch correlation coefficients β _{1 to} β ₃ based on the equation (10) while obtaining the autocorrelation coefficient, and quantizes / decodes the pitch prediction coefficients β _{1 to} β _3. Supply to the chemicalizer 11.

The quantizer / decoder 11 encodes β _{1 to} β ₃ and multiplexes them.
22 and decodes it to pitch predictor
Supply to 6,21. This encoding / decoding is performed in order to make the quantization error conditions the same as those on the combining side.

The multi-pulse train output from the multi-pulse preliminary analyzer 12 has its amplitude quantized by the pulse amplitude quantizer 13 and the memory 14
Stored in.

In this way, the perceptual weighted voice input is stored in the memory 8.
A pitch prediction residual waveform is stored in 10, and an amplitude quantized multipulse train is stored in the memory 14.

The DP processor 15 uses the necessary data stored in each of these memories and uses the pulse for each analysis frame that constitutes the preliminary tentatively determined amplitude-quantized multi-pulse train to obtain the optimum multi-pulse train by the DP method. Determine as follows.

The number of pulses of the tentatively determined multi-pulse train output from the multi-pulse preliminary analyzer 12 is larger than the required number for the purpose of applying the DP method. For example, if the required number is 16 in 4 bits for each analysis frame, this increase is set to about 20.

The processing to be determined by the DP processor 15 is to make a final determination of the optimum pulse from which extra pulses have been removed from the tentatively determined multi-pulse, and a determination of the optimum quantization of the finally determined pulse. DP that minimizes the error power between the speech waveform and the reproduced speech waveform
It is done in the form of finding a path.

DP processing that removes unnecessary pulses and makes the optimum pulse necessary
Taking the case of this embodiment as an example, the operation is basically performed as follows.

That is, the candidate pulse that can be the first in the final determination among the 20 pulses that are temporarily determined in each analysis frame can be selected only from the first to fifth pulses of the temporary determination pulse. This is based on the fact that if the sixth pulse is used as the final determination pulse, five pulses will be removed.

Subsequently, the pulse that can be the second candidate for the final determination can be selected from the tentatively determined first to sixth pulses, except for the pulse that is the first pulse for the final determination. With such a combination, a plurality of selectable pulse trains consisting of 20 pulses are selected, an audio signal is reproduced for each pulse train, and a multi-pulse that forms a pulse train that minimizes the error power between these reproduced voices is generated. This is determined as the optimum multi-pulse train. Therefore, the optimum multi-pulse train can be determined only by evaluating the error power corresponding to the selectable pulse train, and the amount of calculation required for the determination can be greatly reduced as compared with the brute force combination evaluation.

On the other hand, the determination of the optimum quantization level is basically performed as follows.

That is, 20 pulses are tentatively quantized once, and a predetermined quantization step such as +1 and -1 lower value is prepared for each pulse, and the three quantization values are obtained for each pulse. The optimum quantization level is determined for each selectable multi-pulse train for each analysis frame in the DP format that selects the optimum level among the levels. The DP processing for selecting the quantization level is also performed using the error power as an evaluation measure.

The above-described DP process for selecting the optimum pulse and determining the optimum quantization level may be either a temporally continuous separation process or a parallel process. It is processing.

In this way, the error power is used as the evaluation scale, and the optimum pulse and its quantization level are determined in the form of selecting the optimum DP path.However, this DP processing method has been introduced in many documents. The DP processing contents similar to those in the case of implementation are also described in detail, for example, in Japanese Patent Application No. 59-128730, multi-pulse type encoding device.

The DP processor 15 supplies a selectable multi-pulse train as a driving sound source of the LPC synthesis filter 17 for each analysis frame next, and the LPC synthesis filter 17 uses γα input from the attenuation coefficient calculator 5 as its filter coefficient to weight the sound. The voice is reproduced and supplied to the adders 18 and 20. DP processor 15
The multi-pulse train supplied from LPC synthesis filter 17 to the LPC synthesis filter 17 is the optimum number of candidates for the multi-pulse train necessary for the DP processor 15 to make a final decision as the optimum multi-pulse train, and the error power is calculated for each candidate. From this background, the DP processing route is represented by a thick arrow. The LPC synthesis filter 17 uses an all-pole digital filter.

Now, the adder 18 outputs the difference between the data regarding the pitch prediction residual waveform received from the memory 10 and the data regarding the reproduced waveform received from the LPC synthesis filter 17, for each DP path,
This is supplied to the error power calculator 19.

The error power calculator 19 is a DP processor while calculating the error power with respect to the weighted audio input waveform based on this input.
It is supplied to 15 as data for evaluation scale.

The pitch predictor 21 forms a pitch prediction filter together with the adder 20, receives the reproduced waveform from the LPC synthesis filter 17, generates a predicted waveform, and supplies this to the adder 16.

The pitch predictor 21 corresponds to the pitch predictor 51 on the analysis side in FIG. 6 described above, functions in the pitch prediction loop, and significantly reduces the quantization noise due to the feedback effect, and is added to the adder 20 and the adder 16. Which provides the predicted waveform,
Relating this to FIG. 6, the data sent from the adder 20 to the pitch predictor 21 corresponds to + y-, and the output from the pitch predictor 21 corresponds to y. Therefore, the adder 16 to the DP processor The input to 15 is a pitch prediction residual waveform represented by xy.

The pitch prediction residual waveform input to the DP processor 15 is always improved with S / N while receiving the feedback effect of such a pitch prediction loop.

The memory 14 is used for storing the update data of the multi-pulse amplitude value by the quantization level optimum DP processing by the DP processor 15.

The optimum multi-pulse position and amplitude information thus obtained are supplied from the DP processor 15 to the multiplexer 22.

The multiplexer 22 receives the K provided from the K quantizer / decoder 30.
The parameters, the pitch prediction coefficient obtained from the quantizer / decoder 11 and the data related to the multi-pulse provided from the DP processor 15 are appropriately combined and multiplexed, and the multiplexed data is sent to the combining side via the transmission path.

Thus, in a state in which the multi-pulse coding portion is inserted and arranged in the pitch prediction loop, and further, the influence of the quantization noise due to the multi-pulse coding system can be reliably mitigated by the feedback effect of the pitch prediction loop. The analysis side can be realized.

Next, the combining side will be described. FIG. 2 is a block diagram showing a first embodiment on the combining side of the multi-pulse encoding apparatus according to the present invention.

In the embodiment shown in FIG. 2, the demultiplexer 23, the multi-pulse decoder 24, the decoder 25, the LPC synthesis filter 26, the attenuation coefficient calculator 27, the K / α converter 28, the adder 29, and the pitch. Decoder 3
0, pitch predictor 31 and H (Z) / H '(Z) converter 32
It is configured with.

The demultiplexer 23 demultiplexes the multiplexed data input transmitted from the analysis side, and multipulses, K parameters, and pitch prediction coefficients are respectively decoded by the multipulse decoder.
24, K decoder 25, and pitch decoder 30.

The multi-pulse decoder 24 decodes the encoded multi-pulse into original data and supplies it as an input to the LPC synthesis filter 26.

The K decoder 25 decodes the input K parameter and supplies it to the K / α converter 28.

The K / α converter 28 converts the input K parameter into an α parameter, which is used by the attenuation coefficient calculator 28 and H (Z) / H ′ (Z).
Supply to the converter 32.

The attenuation coefficient calculator 27 multiplies the input α parameter by the attenuation coefficient γ and sends γα to the LPC synthesis filter 26. The value of γ in this case is the same as that on the analysis side and is set in advance. The LPC synthesis filter 26 receiving this γα uses this as a filter coefficient to form an all-pole digital filter having a transfer relationship H '(Z).

The LPC synthesizing filter 26 is driven with the multi-pulse as an input, outputs a reproduced voice and supplies it to the adder 29.

The pitch decoder 30 decodes the pitch prediction coefficient and supplies it to the pitch predictor 31. The pitch predictor 31 forms a pitch prediction filter together with the adder 29, inputs the reproduced waveform output from the adder 29, and uses this and the pitch prediction coefficient to output the predicted waveform from the next adder 28. , This is H
The (Z) / H '(Z) converter 32 is supplied.

The H (Z) / H '(Z) converter 32 has a characteristic opposite to that of the H' (Z) / H (Z) converter 1 which is the perceptual weighting filter shown in FIG. Therefore, the reproduced waveform is released as an audio output after the addition of the attenuation constant for weighting the listening sound is canceled. The α parameter and the γα parameter necessary for forming the filter of the H (Z) / H ′ (Z) converter 31 are supplied from the K / γ converter 25 and the attenuation coefficient calculator 27, respectively.

FIG. 3 is a block diagram showing a second embodiment on the combining side of the multi-pulse encoder according to the present invention. The combining side shown in FIG. 3 includes a demultiplexer 23, a multi-pulse decoder 24,
K decoder 25, K / α converter 28, adder 29, pitch predictor
30, a pitch decoder 30, a pitch predictor 31, an LPC synthesis filter 33, etc., and those having the same symbol numbers in these constituent elements are the same as those shown in FIG. Detailed description will be given in detail.

Compared to the composition side of FIG. 2, the composition side of FIG. 3 does not include the attenuation coefficient calculator 27 and the H (Z) / H ′ (Z) converter 32, and the transmission of the LPC composition filter 33 is different. The difference is that the function is H (Z). This means that the LPC synthesis filter 26 shown in FIG.
The transfer function of is H '(Z), and this and H (Z) / H'
The total filter characteristic of the (Z) converter 32 can be represented by the transfer function of H (Z), and therefore H (Z) is eventually obtained.
Can be replaced by the LPC synthesis filter 33 having the transfer characteristic of. In this connection, the attenuation coefficient calculator 27 is not necessary, and the configuration conversion is performed. With such a configuration, the voice input can be reproduced as in the case of FIG.

〔The invention's effect〕

As described above, the present invention provides a multi-pulse encoding device having a pitch prediction loop, which includes means for determining an optimum multi-pulse train based on pitch prediction for a multi-pulse train that is tentatively preliminarily determined. There is an effect that noise can be effectively and remarkably reduced.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the analysis side of the multi-pulse encoding apparatus of the present invention, and FIG. 2 is a block diagram showing a first embodiment of the composite measurement of the multi-pulse encoding apparatus of the present invention. FIG. 3 is a block diagram showing a second embodiment on the combining side of the multi-pulse encoding device of the present invention, and FIG. 4 is a block showing a multi-pulse code decoding system in which a pitch prediction system and a multi-pulse processing system are independent of each other. FIG. 5 is a block diagram showing a multi-pulse code decoding system in which a pitch prediction system is cascade-connected to a proximity prediction system, and FIG. 6 is a multi-pulse code decoding system in which a multi-pulse coding system is inserted in a pitch prediction loop. FIG. 7 is a block diagram showing the general configuration of the pitch predictor, and FIG. 8 is a basic block diagram of multi-pulse search by correlation region processing. 1 ... H '(Z) / H (Z) converter, 2 ... LPC analyzer,
3 ... K quantizer / decoder, 4 ... K / α converter, 5 ...
Attenuation coefficient calculator, 6 ... pitch predictor, 7 ... pitch prediction coefficient calculator, 8 ... memory, 9 ... adder, 10 ... memory, 11 ... quantizer / decoder, 12 ... Multi-pulse preliminary analyzer, 13 ... Pulse amplitude quantizer, 14 ... Memory,
15 …… DP processor, 16 …… adder, 17 …… LPC synthesis filter, 18 …… adder, 19 …… error power calculator, 20 ……
Adder, 21 …… Pitch predictor, 22 …… Multiplexer, 23 ……
Multiplexer, 24 …… Multipulse decoder, 25 …… K
Decoder, 26 LPC synthesis filter, 27 attenuation coefficient calculator, 28 K / α converter, 29 adder, 30 pitch decoder, 31 pitch predictor, 32 …… H (Z) / H ′
(Z) converter, 33 ... LPC synthesis filter, 34 ... pitch predictor, 35 ... adder, 36 ... multi-pulse encoder,
37 ... Multi-pulse decoder, 38 ... Adder, 39 ... Pitch predictor, 40 ... Multi-pulse encoder, 41 ... Pitch predictor, 42 ... Proximity prediction filter, 43 ... Adder , 44
…… Error analyzer, 45 …… Pitch predictor, 46 …… Proximity prediction filter, 47 …… Adder, 48 …… Multi-pulse encoder, 49 …… Multi-pulse decoder, 50 …… Adder, 51 ...
… Pitch predictor, 52 …… Multi-pulse decoder, 53 ……
Adder, 54 ... Pitch predictor, 55 ... Variable delay circuit, 56
-1,56-2 ... unit delay element, 57-1, 57-3 ... multiplier, 58 ... adder, 59 ... impulse response calculator, 60 ...
… Autocorrelation coefficient calculator, 61 …… Cross-correlation coefficient calculator, 62
...... Multi-pulse searcher.

Claims (1)

[Claims] 1. A means for obtaining a pitch prediction coefficient from an input speech signal, a means for removing a pitch prediction component from an input speech signal, a means for tentatively determining a multipulse from a speech waveform from which the pitch prediction component has been removed, A combination of multi-pulses having the best match with the input speech while reproducing the speech waveform, having a synthesis filter and a pitch predictor in order to perform optimum pulse selection and quantization on a part of the determined multi-pulses. And a multi-pulse encoding device.