JPH08129399A - Voice coding device - Google Patents

Abstract

PURPOSE: To provide a voice coding device in which voice coding is performed employing a smaller amount of calculations. CONSTITUTION: An adding circuit 105 adds an input signal s(i) and an output signal βg(i) of a gain circuit 120. A two input buffer circuit 110 accumulates the output signals s(i)+ βg(i) of the circuit 105 and the output signal g(i) of a simplified interpolating circuit 115. The circuit 115 interpolates the signals accumulated in the circuit 110 employing a prescribed equation and computes a long term prediction signal g(i). The circuit 120 computes the product of the signal g(i) and a long term prediction filter gain β. An output terminal 125 outputs an output signal b(i) of the circuit 105. In the circuit 110, only the signal b(i), (i=n-M to n-1) of the long term prediction filter is accumulated at time n which is the top of a current subframe. After that, the signal b(i) is interpolated using a prescribed equation at every time and the signal g(i) is computed.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech coder,
In particular, the present invention relates to a voice encoding device for encoding a voice signal with high quality at a low bit rate of 4 kbps or less.

[0002]

2. Description of the Related Art A conventional speech coder has conventionally been known as an effective method for coding a speech signal at a low bit rate.
The CELP (Cod Excited Linear Prediction Coding) method is known. Regarding the CELP method, for example,
IEE Proceedings (IEEE Proc.) IC
ASSP-85, 1985, pages 937 to 940 (Reference 1). The CELP method represents an audio signal by its spectral characteristics and its excitation signal. The spectral characteristics are represented by two linear prediction filters, and the short-term prediction filter coefficient, the long-term prediction filter coefficient, and the delay are encoded. The excitation signal is coded by vector quantization using the excitation codebook. The sound source codebook is created in advance using voice data or white noise.

In the CELP method, the short-term prediction filter coefficient of input speech is quantized in frame units created by cutting out into a predetermined length. In addition, the coefficients and delays of the long-term prediction filter and the excitation signal are quantized in units of subframes obtained by further dividing this frame. The output response of the long-term prediction filter is expressed in z-transform notation as follows.

[0004]

The delay L has a value close to the pitch period of the input voice or an integral multiple of the pitch period, or an integral fraction. β is a long-term prediction gain. Also, the output response H (z) of the short-term prediction filter for expressing the outline of the spectral structure of speech is z
It is expressed by the following equation using the conversion display.

[0006]

Α (i) (i = 1, ..., P) is a short-term prediction filter coefficient obtained by performing a linear prediction analysis on the input speech. p is the predicted order.

Next, the updating of the contents of the buffer and the operation of the filter of the conventional long-term prediction filter for realizing the equation (1) will be described with reference to FIG. n is the time of the first sample of the current subframe. b (i), (i = nM, ..., n-1) are past output signals of the long-term prediction filter. M is the maximum delay value, and a value of 120 to 150 is used. The output b (i) of the long-term prediction filter is expressed by the following equation using the long-term prediction signal g (i) and the sound source signal x (i).

B (i) = βg (i) + x (i) (3) The long-term predicted signal g (i) is calculated by the following equation.

[0010]

Here, Q is an integer part of the delay L. ε
(m, R) and (m = -m ₁ , ..., M ₂ ) are coefficients of an interpolation filter that delays the signal by a decimal value R of 1 or less. U = Q + R. A typical method for interpolation filters is Multirate Digital Signa processing by RE Crochiere and LR Rabiner.
l Processing), Chapter 4, Prentice Hall (Prentic
e-Hall), 1983. There is a method using the interpolation finite impulse response filter described in (Reference 2). Further, the buffer signal h (i) of the filter is expressed by the following equation.

H (i) = b (i), i = nM, ..., n-1 (5) When the delay L is longer than the subframe length N, the equation (4) is set to the time of the beginning of the current subframe. By repeating from i = n to i = n + N-1, the long term prediction filter for the current subframe can be performed.

If the delay L is shorter than the subframe length N,
After time n + L, in order to execute equation (4), time n
The subsequent buffer signal h (i) is required. However, if the processing of the current subframe is not completed, the long-term predicted signals b (i), (i = n, ..., n + N-1) of the current subframe are not calculated, and the buffer signal h (i) is not calculated. I can't get it. For this reason,
There is no buffer signal h (i) after time n as in (a). Therefore, it is necessary to approximately create the buffer signals h (i), (i = n, ..., n + N-1).

In the conventional speech coding apparatus, a long-term predicted signal g (i) is created by repeating the existing signal before time n-1 as shown in FIG. There is.

[0015]

This is considered to be a suitable approximation because the waveforms of substantially the same shape are repeated when the pitch structure of the voice signal for a short time is viewed. This operation is represented by the following expression when expressed by the output response of z conversion display.

[0017]

Here, the argument i of p (i, z) is the time in the current subframe. "(I + L) / L" represents the nearest integer less than or equal to (i + L) / L. The delay of the conventional long-term prediction filter changes as a function of time i in the current subframe.

As a method used in a conventional speech coding apparatus, for example, a method described in Japanese Patent Laid-Open Publication No. 4-502675, "Digital Speech Coder with Improved Long Term Predictor" (Reference 3). There is.

FIG. 3 is a block diagram showing the configuration of the present invention and the conventional speech encoding device, and FIG. 6 is a block diagram illustrating the configuration of the long-term prediction filter used in the conventional speech encoding device.

FIG. 3 is a block diagram showing the entire configuration used both for the description of the speech coding apparatus of the present invention and for the description of the conventional speech coding apparatus. First, the input voice signal s (i) is input from the input terminal 10. The coefficient analysis circuit 15 analyzes the input signal for each predetermined frame length. In this analysis, the short-term prediction filter coefficient α (i), the long-term prediction filter coefficient β, the long-term prediction filter delay L, the weighting filter W, and the sound source gain coefficient γ are calculated. A typical method of calculating these parameters is IEEE Transactions on Communications (IEEE Transactions
on Communications), Volume COM-30, 1982, April,
See pages 600-614 (reference 4). The subtractor 20 also calculates a difference signal between the input voice and the reproduced voice signal output from the short-term prediction filter circuit 40. The sound source codebook circuit 25 stores sound source code vectors of length N created in advance. Further, in accordance with the instruction from the codebook search control circuit 55, the sound source code vectors are output to the gain circuit 30 one by one. The gain circuit 30 integrates the input excitation code vector with the coefficient γ output from the coefficient analysis circuit 15, and calculates the coefficient γ by using the long-term prediction filter circuit (35 used in the speech encoding apparatus of the present invention, the conventional one). It is output to 70). The long-term prediction filter circuit 70 uses the sound source code vector in which the coefficient γ is integrated, the coefficient β output from the coefficient analysis circuit 15, and the delay L to execute the calculation. The operation of the long-term prediction filter circuit 70 will be described later. The short-term prediction filter circuit 40 executes the operation of the short-term prediction filter 40 of the equation (2) on the output signal of the long-term prediction filter circuit 70, and outputs the output to the difference circuit 20. At this time, the short-term prediction filter coefficient α (i) output from the coefficient analysis circuit 15 is used. The weight circuit 45 uses the weight function w output from the coefficient analysis circuit 15 to calculate the difference circuit 20.
The weighting of the differential signal output by is performed. Here, as a typical weight function w, there is an auditory filter that emphasizes frequencies that are perceptually more important to the human ear and attenuates other frequencies. The sum of squares circuit 50 calculates the energy of the difference signal weighted by the weighting circuit 45, and outputs this energy to the codebook search control circuit 55. The codebook search control circuit 55 compares the energy of the difference signal calculated for the sound source vector of interest with the energy of the difference signal similarly calculated for the other sound source vector to obtain the minimum energy. The index I of the sound source code vector to have is determined. When the comparison is completed, a command for outputting the next sound source code vector is output to the sound source codebook circuit 25. The multiplexer 60 includes a short-term prediction filter coefficient α (i) calculated by the coefficient analysis circuit 15, a long-term prediction filter coefficient β and a delay L,
The weighting filter w, the excitation gain coefficient γ, and the index I of the excitation code vector determined by the codebook search control circuit 55 are converted into binary data. The output terminal 65 outputs the binary data I to the channel.

Next, the configuration of the long-term prediction filter circuit 70 used in the conventional speech coding apparatus and the signal flow of its buffer circuit 210 will be described with reference to FIGS. 6 and 7. The input terminal 200 is the output signal s (i) of the gain circuit 30 of FIG.
Enter The adder circuit 205 performs addition calculation of the input signal s (i) and the output signal βg (i) of the gain circuit 220. The buffer circuit 210 uses the signal h (i) = s (i) + βg obtained by the addition.
Accumulate (i). The interpolation circuit 215 is a buffer circuit 21.
The long-term predicted signal g (i) is calculated by performing the interpolation of the equation (4) on the signal accumulated in 0. The gain circuit 220 calculates an integrated value of the signal g (i) and the long-term prediction filter gain β. The output terminal 225 outputs the output signal b (i) of the adder circuit 205 to the short-term prediction filter circuit 40 of FIG. As shown in FIG. 7 (a), the buffer circuit 210 displays the signal b at the time n at the beginning of the current subframe.
Only (i) (i = nM ~ n-1) is accumulated. Time n to n
Up to + L, the signal g (i) is calculated by interpolating the signal b (i). Since the excitation signal from time n to n + N-1 cannot be used, after time n + L + 1, the delay L is delayed as shown in Fig. 7 (c).
The long-term predicted signal g (i) is calculated by interpolating the signal h (i-kL) having a delay kL that is an integer multiple of.

Next, the operation of the long-term prediction filter circuit in the conventional speech coder will be described. FIG. 8 is a flow chart of the operation of the conventional long-term prediction filter circuit.

First, when starting at step (hereinafter referred to as S) 500, a sample number i is initialized to n at S505. In S510, the delay L and the coefficient β of the long-term prediction filter 70 are set. Next, in S515, the sample number i is n + N
If it is larger than -1, the operation is ended in S555, and if it is not larger, the signal s (i) is input in S520. In step S525, the signal g (i) is calculated by equation (6), and in step S530, the output response b (i) of the long-term prediction filter 70 is calculated using equation (3). Subsequently, in S535, the buffer signal h (i) is updated by the equation (11), and in S540, the output sample b (i) is input to the last h (-1) of the buffer. In S545, output sample b (i)
Is output. Finally, in S550, the sample number i is updated, and the process returns to S505 to perform the calculation again.

[0025]

In the conventional speech coding apparatus described above, since the filter processing of the long-term prediction filter 70 to be used changes as a function of the pitch period L and the time i as shown in equation (6), There is a problem that the filtering process becomes complicated and the amount of calculation increases.

An object of the present invention is to solve the above-mentioned problems and to provide a speech coding apparatus which performs speech coding with a smaller amount of calculation.

[0027]

A speech coding apparatus according to a first invention of the present application comprises a long-term prediction filter circuit for coding a pitch period of a decimal point value of an input speech signal, and the input speech is defined in a predetermined subframe. In a speech coding apparatus for coding for each length, the long-term prediction filter receives a buffer signal representing the state of its present output value, performs an interpolation calculation on this buffer signal, and outputs an interpolation value which is a calculation result. An interpolation circuit, a feedback gain circuit that calculates a feedback value by integrating a long-term predicted gain with the interpolation value output by the interpolation circuit, a feedback value calculated by the feedback gain circuit, and an input value of a newly arrived subframe. And a buffer circuit including an output value output from the addition circuit and an interpolation value output from the interpolation circuit. A configuration and a buffer circuit for storing.

A speech coder according to the second invention of the present application comprises a long-term prediction filter circuit for coding the pitch period of the decimal point value of the input speech signal, and encodes the input speech for each predetermined subframe length. In the speech coding device, the long-term prediction filter receives a buffer signal representing a state of a current output value of the self, performs an interpolation calculation on the buffer signal, and outputs an interpolation value which is a calculation result, and the interpolation circuit. A gain circuit for calculating an integration result value by integrating a feedback gain with an interpolation value output by the circuit; and a feedback gain circuit for calculating a feedback value by integrating a long-term predicted gain with the interpolation value output by the interpolation circuit. An adder circuit for adding the feedback value calculated by the feedback gain circuit and an input value of a newly arriving subframe to output as a new output value, A configuration and a buffer circuit for storing the buffer signal and an output to the output value of the serial adder circuit and the integration result value output to the gain circuit.

[0029]

According to the present invention, when the delay L of the long-term prediction filter is shorter than the subframe length N, the approximate value of the output signal of the long-term prediction filter of the current subframe, which is necessary for executing the equation (1), is set to By generating the samples one by one while executing the calculation, the long-term prediction filter can always be executed by the same filtering process.

In the first invention, the long-term predicted signal g (i) which is the predicted value of the input signal is used as the approximate value of the output signal of the filter after time n. The long-term predicted signal g (i) is created by the following equation.

[0031]

Here, the updating of the contents of the buffer of the long-term prediction filter 35 and the operation of the filter of the speech encoding apparatus of the first invention will be described. Let n be the time of the first sample of the current subframe.

B (i), (i = nM, ..., n-1) are past output signals of the long-term prediction filter. From time n to n + L, past output signals b (i), i = nM, ..., n-1 of the long-term prediction filter
Is used to calculate the long-term predicted signal g (i) by the equation (8). At this time, the long-term predicted signal g (i) is stored in the buffer signal h (i) at each time as shown in the equation (9). After time n + L + 1, the buffer signal h (i) is used, and according to equation (8),
Calculate the long-term predicted signal g (i). Where g (i) is replaced by h (i)
If they are stored in the buffer circuit of, the filter processing can be executed by the simple processing of the expression (8) without performing the complicated processing of the conventional method of the expression (6).

Further, in the second invention, the long-term prediction filter 3
In order to improve the approximation accuracy of the long-term predicted signal g (i) with respect to the output signal of No. 5, the following equation in which the gain η is integrated is used. As the value of the gain η, the value of β of the previous subframe can be used.

[0035]

As described above, the long-term prediction filter in the speech coding apparatus can be performed by simple and identical filter processing instead of complicated filter processing that has been conventionally performed, and hardware such as DSP is used. Not only can the simplification process be simplified, but the amount of calculation can be reduced.

[0037]

Embodiments of the present invention will now be described with reference to the drawings.

FIG. 1 is a block diagram of a long-term prediction filter circuit of a speech coder according to a first embodiment of the present invention, and FIG. 2 is a long-term prediction filter circuit of a speech coder according to a second embodiment of the present invention. FIG. 3 is a block diagram, FIG. 3 is a block diagram showing an overall configuration of the present invention and a conventional speech coding apparatus, and FIG.

The description of the operation of the entire speech coding apparatus is the same as that described in the section of the prior art and is redundant, and therefore the description thereof is omitted here. However, names and reference numerals are used in the description as appropriate.

First, the configuration of the long-term prediction filter circuit 35 in the first invention will be described with reference to FIGS. 1 and 4.

The output signal s (i) of the gain circuit 30 (FIG. 3) is input to the input terminal 100. The adder circuit 105
Input signal s (i) and output signal βg (i) of gain circuit 120
The addition calculation of is performed. The 2-input buffer circuit 110 stores the output signal s (i) + βg (i) of the adder circuit 120 and the output signal g (i) of the simple interpolation circuit 115. The simple interpolation circuit 115 interpolates the signal stored in the 2-input buffer circuit 110 by the equation (8) to calculate the long-term predicted signal g.
Calculate (i). The gain circuit 120 calculates an integrated value of the signal g (i) and the long-term prediction filter gain β. The output terminal 125 outputs the output signal b (i) of the adder circuit 105 to the short-term prediction filter circuit 40 (FIG. 3). As shown in FIG. 1, the 2-input buffer circuit 110 stores only the output signals b (i) and (i = nM to n-1) of the long-term prediction filter at the time n at the beginning of the current subframe. ing. From time n to n + L, the signal g (i) is calculated by interpolating the signal b (i) using the equation (8). After time n + L + 1,
Compute the signal g (i) by interpolating the signal g (i) using equation (8). The above-described interpolation calculation is for generating an excitation signal corresponding to the decimal point delay L.

Next, the configuration of the long-term prediction filter circuit 35 in the second invention will be described with reference to FIG.

The output signal s (i) of the gain circuit 30 (FIG. 3) is input to the input terminal 100. The adder circuit 105
Input signal s (i) and output signal βg (i) of gain circuit 120
The addition calculation of is performed. The 2-input buffer circuit 110 includes a past output signal ηg (i) of the gain circuit 130 and the addition circuit 10
5 output signals s (i) + βg (i) are accumulated. The simple interpolation circuit 115 calculates the long-term predicted signal g (i) by interpolating the signal stored in the two-input buffer circuit 110 by the equation (8). The gain circuit 120 outputs the signal g (i)
And the integrated value of the long-term prediction filter gain β is calculated. The gain circuit 130 calculates the integrated value of the signal g (i) and the long-term prediction filter gain η. The output terminal 125 outputs the output signal b of the addition circuit 105 to the short-term prediction filter circuit 40 (FIG. 3).
Output (i).

Next, the operation of the long-term prediction filter circuit 35 of the present invention will be described. FIG. 5 is a flow chart of the operation of the long-term prediction filter circuit 35 of the present invention.

First, when starting at S400, the sample number i is initialized at time n at the beginning of the current subframe at S405. Next, in S410, the delay L as a long-term prediction filter and the coefficient β are set, and in S415, the sample number i
Is larger than the last time n + N-1 of the current subframe, and if larger, the operation ends in S455. If not, input the signal s (i) in S420,
At 425, the signal g (i) is calculated using equation (8). S43
At 0, the output response b (i) of the long-term prediction filter is calculated by using the equation (3), and then at S435, the buffer signal h (i) is updated by the following equation. h (i) = h (i + 1), i = -M to -2 (11) Next, in S440, the output sample g (i) previously output in S425 is added to the last h (-1) of the buffer signal. ), And press S4
The output sample b (i) is output at 45, the sample number i is updated at S450, and the process returns to S415 to perform the calculation again.

Note that the long-term prediction filter coefficient β and the gain coefficient γ of the sound source codebook are L in the closed loop configuration.
Can be optimized simultaneously for all values of.

Also, the delay L of the long-term prediction filter and the coefficient β
When determining the index of the excitation code vector and the coefficient γ in order, when determining the delay L and the coefficient β of the long-term prediction filter, the excitation vector of the current subframe is undecided and cannot be used. If the delay L is smaller than the subframe length N, the filter of equation (1) cannot be realized. In this case, the long-term prediction filter can also assume that all the elements of the excitation code vector of the current subframe are 0.

Further, the order of the long-term prediction filter may be second or higher.

Further, the long-term prediction filter in the present invention is
Any form of speech coding (eg RELP, multipulse, RPE, LPC) can also be used.

FIG. 9 is a block diagram showing an embodiment of a voice decoding device for decoding a voice signal using binary data output from the voice encoding device of the present invention. The demultiplexer 300 uses the binary data input from the input terminal 305 to calculate the short-term prediction filter coefficient α (i), the long-term prediction filter delay L and the coefficient β, the excitation code vector index I, and the gain coefficient γ. The sound source codebook circuit 310 outputs a sound source code vector of index I. The gain circuit 320 integrates the input sound source code vector with the gain coefficient γ. The long-term prediction filter circuit 330 performs the same operation as the long-term prediction filter circuit 35 of the speech encoding apparatus of the present invention, and thus the delay L and the coefficient β
Is used to perform long-term predictive filtering. The short-term prediction filter circuit 340 outputs, from the output terminal 350, a reproduced audio signal created by short-term prediction filtering the output signal of the long-term prediction filter circuit 330 using the short-term prediction filter coefficient α (i).

[0051]

As described above, according to the present invention, the long-term prediction filter receives a buffer signal representing the state of its own current output value, performs an interpolation calculation on this buffer signal, and calculates the interpolation value as the calculation result. The output interpolation circuit, the feedback gain circuit that calculates the feedback value by adding the long-term predicted gain to the interpolation value output by the interpolation circuit, the feedback value calculated by the feedback gain circuit, and the newly arrived subframe input value. And a buffer circuit for accumulating a buffer signal composed of the output value output by the adder circuit and the interpolated value output by the interpolator There is an effect that the audio coding can be performed by the amount.

[Brief description of drawings]

FIG. 1 is a block diagram of a long-term prediction filter circuit of a speech coder according to a first embodiment of the present invention.

FIG. 2 is a block diagram of a long-term prediction filter circuit of a speech coder according to a second embodiment of the present invention.

FIG. 3 is a block diagram showing an overall configuration of a speech encoding apparatus of the present invention and a conventional speech encoding apparatus.

FIG. 4 is an explanatory diagram illustrating an operation of a signal of a buffer circuit of a long-term prediction filter.

FIG. 5 is a flow chart of the operation of the long-term prediction filter circuit 35 of the present invention.

FIG. 6 is a block diagram illustrating a configuration of a long-term prediction filter used in a conventional speech encoding device.

FIG. 7 is an explanatory diagram illustrating a signal operation of a conventional buffer circuit.

FIG. 8 is a flow chart of the operation of a conventional long-term prediction filter circuit.

FIG. 9 is a block diagram showing an embodiment of a speech decoding apparatus that decodes a speech signal using binary data output from the speech encoding apparatus of the present invention.

[Explanation of symbols]

 10, 100 Input terminal 15 Coefficient analysis circuit 20 Subtractor 25 Sound source codebook circuit 30, 120, 130 Gain circuit 35 Long-term prediction filter circuit 40 Short-term prediction filter circuit 45 Weighting circuit 50 Square sum circuit 55 Codebook search control circuit 60 Multiplexer 65 , 125 output terminal 105 adder circuit 110 2 input buffer circuit 115 simple interpolation circuit

Claims (2)

[Claims] 1. A speech coding apparatus, comprising: a long-term prediction filter circuit for coding a pitch period of a decimal point value of an input speech signal, wherein the input speech is coded for each predetermined subframe length. Is an interpolator that receives a buffer signal representing the current output value state of itself and performs an interpolation operation on this buffer signal and outputs an interpolated value that is the operation result, and a long-term prediction gain for the interpolated value output by the interpolator. A feedback gain circuit that calculates a feedback value by accumulating, and an addition circuit that adds the feedback value calculated by the feedback gain circuit and the input value of a newly arriving subframe and outputs the result as a new output value, A buffer circuit for accumulating a buffer signal composed of the output value output from the adder circuit and the interpolated value output from the interpolator. Speech coding device. 2. A speech coding apparatus comprising a long-term prediction filter circuit for coding a pitch period of a decimal point value of an input speech signal, wherein the input speech is coded for each predetermined sub-frame length. Receives a buffer signal representing the current output value state of its own, performs an interpolation operation on this buffer signal, and outputs an interpolation value as the operation result, and a feedback gain to the interpolation value output by the interpolation circuit. A gain circuit for calculating an integrated result value by integrating, a feedback gain circuit for calculating a feedback value by integrating a long-term predicted gain with the interpolation value output by the interpolation circuit, and a feedback value calculated by the feedback gain circuit. And an input value of a newly arrived subframe and output as a new output value; an output value output by the addition circuit; And a buffer circuit for accumulating a buffer signal composed of an integration result value output from the gain circuit.