JP3462958B2 - Audio encoding device and recording medium - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice encoding device and a recording medium recording a program that implements the same with software.

[0002]

2. Description of the Related Art As an 8 kbit / s voice encoding system, IT
Conjugate-Structure Algebr standardized by UT
aic-Code-Excited-Linear-Predictive (CS-ACELP) Coding
(Draft Recommendation G. 729) is known. FIG. 6 shows a block diagram of CS-ACELP. In FIG. 6, 31 is a buffer memory for buffering an input signal, 32 is an LPC analysis circuit for performing LPC analysis of the input signal, 33 is a quantizer for quantizing LPC coefficients, and 34 is for reproducing synthesized speech from a sound source signal. A synthesis filter, 35 is an adder for subtracting a synthetic speech from an input speech signal to obtain a residual signal, 36 is a weighting circuit for perceptually weighting the obtained residual signal, and 37 is an adaptive codebook 38 for storing past driving sound sources. An adaptive codebook search circuit for searching an adaptive code vector, 39 is a fixed codebook search circuit for searching a fixed code vector from a fixed code book 40 that stores a fixed sound source vector such as a noise sound source, and 41 is a gain code book 42. The gain quantizer for predicting the gain, and 43 are quantized LPC coefficients and searched respectively Dobekutoru and a quantization gain a multiplexer for coding and multiplexing. This method has a frame length of 8
0 dimension (10ms), subframe length 40 dimension (5
Sound source codebooks 38 and 40 are searched for each ms).
17 bits are assigned to the sound source information, and 17 bits are configured to represent the positions and codes of the four sound source pulses.

[0003]

[Problems to be Solved by the Invention] This CS-ACELP
In order to further reduce the bit rate (4 kbit / s), the frame length is set to 160 dimensions (20 ms), the subframe length is set to 80 dimensions (10 ms), and the other configurations are exactly the same as in the case of 8 kbit / s, The problem is that since only four sound source pulses can be searched and quantized in one subframe, there is a limit in faithfully reproducing the input voice.

Therefore, in a low bit rate voice encoding system of about 4 kbit / s, a voice encoding system which faithfully reproduces an input voice at a bit rate as low as possible is required.

The present invention provides a relatively low bit rate,
The purpose is to construct a sound source code vector that is faithful to the input speech.

[0006]

In order to achieve the above-mentioned object, the present invention utilizes the redundancy of the residual signal in the vicinity of the pitch pulse to make the relative distance from the pitch pulse centered on the pitch pulse. A pitch adaptive excitation code vector is obtained by a learning method that takes into consideration the above, and a pitch adaptive excitation codebook is created by this method, so that when the speech is encoded at a low bit rate of about 4 kbit / s, the bit rate is as low as possible. The sound source codebook is constructed so as to construct a sound source code vector faithful to the input speech.

Further, by dividing the pitch adaptive excitation codebook into a plurality of pieces, the memory amount and the calculation amount are smaller than the size of the conventional CELP type excitation codebook, and the conventional CS-ACELP is 4 kbit / s. It is possible to obtain a sound source code vector that faithfully reproduces the input voice, rather than the sound source code vector when it is converted.

[0008]

BEST MODE FOR CARRYING OUT THE INVENTION In the present invention , a buffer memory for buffering an input signal, an LPC analysis circuit for performing LPC analysis of the input signal, a quantizer for quantizing LPC coefficients, and a synthesized voice from a sound source signal. A synthesis filter for reproduction, a weighting circuit for perceptually weighting a residual signal obtained by subtracting the synthetic speech from the input speech, a pitch period is predicted from the weighted residual signal, and an adaptive code vector is searched using an adaptive codebook. An adaptive codebook search circuit,
A pitch adaptive noise source for generating and quantizing a noise source adapted to pitch, a gain quantizer and gain codebook for predicting gain, and a multiplexer for multiplexing each quantization parameter are provided . Noise source is sound source code
Divide the book and the code vector of each codebook is
It is orthogonal to the code vectors of other codebooks.
Part of the above code vector has its origin at the pitch position.
Learn in the negative direction of the time axis, and some are positive
It is a code vector that has been learned toward the
Shift the chord vector to the pitch position and add
And cut off the part that exceeds the subframe length.
Provided is a speech encoding device characterized by making .

In the speech coding apparatus of the present invention, the pitch pulse
By using the redundancy of the residual signal near the
Voice bit faithful to the input voice at a low bit rate of it / s
It has the effect that encryption can be performed. As the configuration of the pitch adaptive noise source, a pitch adaptive sound source codebook in which a pitch position is fixed and noise pulses in the vicinity of the noise pulse are learned using the relative distance from the pitch position as the horizontal axis, and the sound source codebook is used. By using a pitch adaptive excitation codebook search circuit that performs a search and generates a pitch adaptive excitation code vector, it is possible to encode speech that is faithful to the input speech at a low bit rate of about 4 kbit / s. Have.

According to a second aspect of the invention, there is provided the speech coding apparatus according to the first aspect, wherein the pitch adaptive excitation codebook search circuit comprises a pitch position detection circuit and a codebook search circuit. By performing a codebook search using the specified pitch position and generating a sound source code vector that is adapted to the pitch, it is possible to encode audio that is faithful to the input audio at a low bit rate of about 4 kbit / s. Have an effect.

A third aspect of the present invention is the speech coding apparatus according to the second aspect , wherein the pitch position detection circuit searches for a position where the amplitude of the adaptive code vector is maximum, and the pitch position is easily searched. It is possible to do so.

According to a fourth aspect of the invention, a pitch adaptive excitation code vector is expressed by a combination of a plurality of n codebooks, and n codebooks are coded so that the codevectors of each codebook are orthogonal to those of other codebooks. 4. The speech coding apparatus according to claim 1, further comprising a codebook in which a pulse is arranged for each and the vector length is compressed to 1 / n times the subframe length, and the codebook is divided into a plurality of codebooks. By performing the search, it is possible to perform the search with a low calculation amount, and by making the vector length 1 / n times the subframe length, it is possible to reduce the memory amount.

According to a fifth aspect of the present invention, a pitch adaptive excitation code in which the codebook search circuit minimizes the mean square error of the weighted residual signal obtained by weighting the residual signal obtained by subtracting the synthetic speech from the input speech. The voice coding device according to claim 2 or 3 , which searches / generates a vector, and shifts the pitch adaptive excitation code vector by the detected pitch position, and searches for an optimum vector from among the 4 kbit / s. It has an effect that it is possible to encode the voice faithful to the input voice at a low bit rate of about s.

According to a sixth aspect of the present invention, the pitch adaptive excitation codebook according to the fourth aspect is used, each codebook is searched by the codebook search circuit according to the fifth aspect , and a plurality of n obtained codes are obtained. by taking the linear sum of vectors, also claim 2 obtain pitch adaptive excitation codevector
Is a voice encoding device described in 3 , and has an effect that it is possible to perform voice encoding faithful to an input voice at a low bit rate of about 4 kbit / s with a low calculation amount and a low memory amount.

The invention described in claim 7 is a recording medium such as a magnetic disk or a magneto-optical disk ROM cartridge which records a program in which the audio encoding device according to any one of claims 1 to 6 is realized by software. By inputting these recording media to, for example, a personal computer or the like, there is an effect that the speech encoding device according to any one of claims 1 to 6 can be realized by software.

FIG. 1 shows an embodiment of the present invention.
4 to FIG. (Embodiment 1) In FIG. 1, 11 is a buffer memory for buffering an input signal, and 12 is an LP of the input signal.
An LPC analysis circuit that performs C analysis, 13 a quantizer that quantizes LPC coefficients, 14 a synthesis filter that reproduces synthesized speech from a sound source signal, and 15 that subtracts synthetic speech from an input speech signal to obtain a residual signal , 16 is a weighting circuit for weighting the residual signal obtained by perceptual weighting, 17 is an adaptive codebook search circuit for searching an adaptive code vector from an adaptive codebook 18 that stores past driving sound sources, and 19 is a pitch adaptive sound source codebook 20. Is a pitch adaptive excitation codebook search circuit that searches for a vector that minimizes the mean square error of the weighted residual signal, and includes a pitch position detection circuit 19A and a codebook search circuit 19B. Reference numeral 21 is a gain quantizer that performs gain prediction using the gain codebook 22, and reference numeral 23 is a multiplexer that multiplexes and codes the quantized LPC coefficient, the searched code vector, and the quantized gain. is there.

Next, the operation of this embodiment will be described. The input signal is buffered in the buffer memory 11 and divided into subframe lengths. The audio signal divided into sub-frame lengths is LPC-analyzed by the LPC analysis circuit 12.
Coefficients are calculated, quantization is performed by the quantizer 13, one of the outputs is input to the multiplexer 23 for coding, and the other is LP
The C coefficient is inversely quantized and input as a coefficient of the synthesis filter 14. The synthesis filter 14 includes a gain quantizer 2
The sum of the adaptive code vector and the pitch adaptive sound source vector, which are respectively scaled by 1, is input, and the synthesized speech is obtained there. A residual signal is obtained by subtracting the synthesized speech from the input speech by the adder 15, and the weighted residual signal is obtained by passing the residual signal through the weighting circuit 16. The adaptive codebook search circuit 17 inputs the weighted residual signal and calculates the pitch period so that the mean square error thereof is minimized. The calculated pitch period is input to the multiplexer 23 and encoded. Then, an adaptive code vector is generated from the adaptive codebook 18 based on this pitch period.

Pitch adaptive excitation codebook search circuit 19
An adaptive code vector and a weighted residual signal are input to. First, a pulse position having the maximum amplitude in the adaptive code vector is searched as a pitch position. Then, using the pitch adaptive sound source codebooks 20 _{1 to} 20 _n that have been learned by taking the relative distance from the pitch position as the horizontal axis, the code vector is shifted to the pitch position and a search is performed. In the search, a vector that minimizes the mean square error of the weighted residual signal is searched in each of the codebooks 20 _{1 to} 20 _n . As a result of the search, n excitation code vectors and their indexes are obtained, and the indexes are input to the multiplexer 23 and encoded. Also, a linear sum of n code vectors is taken to generate a final pitch adaptive excitation code vector. An adaptive code vector, a pitch adaptive excitation code vector, and a weighted residual signal are input to the gain quantizer 21, and the adaptive code vector and the pitch adaptive are adjusted so that the mean square error of the weighted residual signal is minimized. Obtain the gain of the sound source code vector. The obtained gain is quantized by the gain codebook 22 and is multiplexed by the multiplexer 23.
, And each vector is scaled and added to generate a sound source signal. A synthetic voice is obtained by passing the sound source signal through the synthesis filter 14.

Next, details of the pitch adaptive excitation codebook search circuit 19 will be described. FIG. 2 is a flow chart showing the processing procedure of the pitch adaptive excitation codebook search circuit 19, and FIG. 2A shows the overall processing flow of the pitch adaptive excitation codebook search circuit 19 in step S.
1 is the operation of the pitch position detection circuit 19A, and the operations of steps S2 to S6 are the codebook search circuit 19B.
Is the operation. FIG. 2B shows a pitch position detection circuit 19A.
This shows the details of the operation of (1) and searches for the position where the amplitude is maximum in the input adaptive code vector.

Regarding the pitch position detection method, a method for detecting a pitch position that minimizes waveform distortion between a pulse train and an adaptive code vector, or a method for detecting it in a synthesized waveform area,
The accuracy of pitch position detection can be improved by using a method of detecting a pulse position that minimizes waveform distortion between a composite waveform of a pulse train and a composite waveform of an adaptive code vector. Also, using the pitch period,
By making the obtained pitch adaptive sound source code vector into a pitch period, it is possible to improve the performance in the voiced sound part.

FIG. 3 shows the processing procedure of the codebook search circuit 19B.
The codebook is shifted to the pitch position and weighted to search for a code vector that minimizes the mean square error with the input speech.

The same operation is performed for the second codebook and thereafter, but in the search after the second codebook, the search is performed by using the vector obtained by adding the code vector determined before that. By the same operation, the optimum code vector is obtained in each codebook, and the sum of them is the final pitch adaptive excitation code vector.

The above processing is illustrated in FIG. 4, and it can be seen that the generated pitch adaptive excitation code vector is represented by the sum of vectors orthogonal to each other.

The code vectors are determined one by one starting from the first codebook. First, a plurality of optimal candidates are listed from each codebook, and the optimal search is performed among the listed candidates. By providing a process for detecting a combination, that is, a preselection process, a more optimal pitch adaptive excitation code vector can be obtained.

FIG. 5 shows how the code vectors shift to the pitch position. Each code vector is
Since learning is done by taking the relative distance from the pitch position as the horizontal axis, if you know the position information of the pitch, shift each code vector to the pitch position, and discard the components that exceed the subframe length. can get.

Further, at this time, by using the following conversion formula for shifting the pulse position at the position i in the memory area to the position i ′ at the time of the search, 1 / n times the subframe length in the memory area is used. The code vector compressed to
At the time of search, the subframe length can be restored.

[0027]

[Equation 1] However, i: pulse position in memory area, i ': pulse position at the time of search, F: subframe length, n: number of divisions of codebook, k: codebook number (0,
1 ,. ． ． n-1), L: pitch position.

FIG. 5 shows an example in which the subframe length is 80 dimensions, the pitch adaptive excitation codebook is divided into 5, and the pitch position is L = 58. In this case, the code vector length in the memory area is 80/5. = 16, and the pulse position (0, 1, ...,
Up to 7), the negative region on the time axis, pulse position (8,
Since 9 ..., 15) is learned as a positive or origin, when converting the code vector at the time of search, the first term i-8 on the right-hand side of the equation (1) gives 8 from the pulse position of the memory area. Is being reduced. Further, in the second term k−1, the orthogonal relationship is guaranteed by shifting the codebook number by 1 respectively, and in the third term, the pitch adaptation is performed by shifting the pitch position L. To be done.

[0029]

As described above, according to the present invention, by utilizing the redundancy of the residual signal in the vicinity of the pitch pulse, for example, the subframe length is 80 dimensions, the excitation codebook is divided into five, and one excitation codebook is used. 3 bits per type (8 types)
By assigning, the sound source code vector adapted to the pitch can be represented by 15 bits per subframe, and when the speech is encoded at a low bit rate of about 4 kbit / s, a low calculation amount and a low memory amount are required. The advantageous effect that a sound source vector faithful to the input voice can be constructed is obtained.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a speech coding apparatus according to an embodiment of the present invention.

FIG. 2A is a flowchart showing a search processing procedure of a pitch adaptive excitation codebook search circuit in the same apparatus; FIG. 2B is a flowchart showing a processing procedure of a pitch position detection circuit in the pitch adaptive excitation codebook search circuit.

FIG. 3 is a flowchart showing a processing procedure of a codebook search circuit in the pitch adaptive excitation codebook search circuit.

FIG. 4 is a list of adaptive excitation code vectors represented by the sum of code vectors.

FIG. 5: Pitch position L of pitch adaptive sound source code vector
Transition diagram showing shifting up to

FIG. 6 is a block diagram showing the configuration of a conventional CS-ACELP speech encoding apparatus.

[Explanation of symbols]

11 buffer memory 12 LPC analysis circuit 13 quantizer 14 synthesis filter 15 adder 16 weighting circuit 17 adaptive codebook search circuit 18 adaptive codebook 19 pitch adaptive excitation codebook search circuit 19A pitch position detection circuit 19B codebook search circuit 20 ₁ ~ 20 _n pitch adaptive excitation codebook 21 gain quantizer 22 gain codebook 23 multiplexer

─────────────────────────────────────────────────── ─── Continuation of front page (58) Fields surveyed (Int.Cl. ⁷ , DB name) G10L 19/12

Claims (7)

(57) [Claims] 1. A buffer memory for buffering an input signal, an LPC analysis circuit for LPC analysis of the input signal, a quantizer for quantizing LPC coefficients, and a synthesis filter for reproducing synthesized speech from a sound source signal. A weighting circuit that performs perceptual weighting on the residual signal obtained by subtracting the synthetic voice from the input voice, and predicts the pitch period from the weighted residual signal,
An adaptive codebook search circuit that searches an adaptive code vector using an adaptive codebook, a pitch adaptive noise source that generates and quantizes a pitch-adapted noise source, a gain quantizer and a gain codebook that perform gain prediction, and , A multiplexer that multiplexes each quantization parameter , wherein the pitch adaptive noise source divides the excitation codebook,
The code vector of each codebook is replaced by the other codebook
Are orthogonal to the code vector of
Part of the kuturu is the negative of the time axis with the pitch position as the origin
The coaches who have learned positively and some of them have learned positively.
The chord vector up to the pitch position.
Shifted and added together, and exceeded the subframe length
A voice coding device characterized by cutting off parts . 2. The speech coding apparatus according to claim 1 , wherein the pitch adaptive excitation codebook search circuit comprises a pitch position detection circuit and a codebook search circuit. 3. The speech coding apparatus according to claim 2 , wherein the pitch position detection circuit searches for a position where the amplitude of the adaptive code vector is maximum. 4. A pitch adaptive excitation codevector is represented by a combination of a plurality of n codebooks, and a pulse is arranged every nth such that the codevector of each codebook is orthogonal to that of another codebook. the speech encoding apparatus according to claim 2 or 3, wherein a codebook obtained by compressing the vector length to 1 / n times the sub-frame length. 5. A codebook search circuit searches and generates a pitch adaptive excitation code vector that minimizes the mean square error of the weighted residual signal obtained by weighting the residual signal obtained by subtracting the synthetic speech from the input speech. The speech encoding device according to claim 2 or 3 . 6. A codebook search circuit according to claim 5 , wherein the pitch adaptive excitation codebook according to claim 4 is used.
4. The speech coding apparatus according to claim 2 , wherein a pitch adaptive excitation code vector is obtained by searching each codebook and taking a linear sum of a plurality of n obtained code vectors. 7. A recording medium recording a program which realizes the voice coding apparatus described software to any one of claims 1 to 6.