JP2702157B2 - Optimal sound source vector search device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Application Field] The present invention compresses information of an audio signal and performs digital transmission,
Alternatively, the present invention relates to an improvement of a speech encoding device that performs accumulation.

[Conventional technology]

Code Excited Linear Prediction (CEL) is a speech coding method that compresses information by separating a speech signal into a parameter representing a synthesis filter and a parameter representing a sound source.
P: Code-Excited Linear Prediction). As an example of CELP, MR Schroeder, BS Atal's "Code Excited Linear Prediction (CELP): High Quality Speech at Very Low Bit Rate" (MR Schroeder, BSAtal, Cod
e-Excited Linear Prediction (CELP): high-quality
speech at very low bit rates, "Proc, IEEE Int.Conf.
Acoust., Speech, Signal Processing, pp. 937-940 (198
5)) (hereinafter referred to as Document 1). In the example shown in Reference 1, a parameter representing a synthesis filter is obtained by analysis every 10 msec, while a parameter representing a sound source corresponding to time divided into voices divided at every 40 points (5 msec when the sampling frequency is 8 KHz). A time series of 40 points of noise generated by random numbers, that is, a 40-dimensional vector (hereinafter, referred to as a sound source vector) is used.

FIG. 2 shows an apparatus for performing the processing performed by the optimum sound source vector search apparatus in Document 1 in the frequency domain. Fig. 2 shows IM Trancoso, BS Atal's “Efficient Procedures for Finding Optimum Innovation Instrumental Coders” (IMTrancoso, BSAtal, “Effici
ent procedures for finding the optimum innovation
in stochastic coders, "Proc.IEEE Int.conf.Acoust., S
peech, Signal Processing, pp.2375-2378 (1986))
FIG. 2 is a diagram showing a conventional optimal sound source vector search device described in (hereinafter, referred to as Document 2). In the figure, 2
Represents the sound source vector, which is a sample value sequence at N points (in the example of Reference 2, N = 40 when the sampling frequency is 8 KHz), is 2 ·
DF obtained by discrete Fourier transform (DFT) at L point (L = 40 in Reference 2)
T sound source vectors (mapped sound source vectors), 1 is the number of M mapped to the L-dimensional distortion evaluation space satisfying the L-dimensional orthogonal condition.
A codebook (mapped excitation codebook) composed of DFT excitation vectors, and 4 is a DFT input speech (DFT input speech) obtained by subjecting an input speech (input speech signal) 11, which is an N-point sample value sequence, to a 2-L point DFT. 5 is an impulse response of a synthetic filter coefficient obtained by analyzing the input speech 11 at 2 · L points.
It is an evaluation weight filter coefficient as a frequency characteristic obtained by DFT. Reference numeral 12 denotes a DFT circuit (first L-dimensional mapping means) for performing DFT on the input voice 11 to map the same into an L-dimensional distortion evaluation space similar to the DFT sound source vector 2;
A speech analysis circuit that analyzes 1 to calculate a synthesis filter coefficient, 15 is an impulse response generation circuit that calculates an impulse response of a synthesis filter coefficient output from the voice analysis circuit 14, and 16 is a speech analysis circuit 14 and an impulse response generation circuit.
Speech analysis means 15 comprises a DFT sound source vector 2 by performing DFT on the impulse response of the synthesis filter coefficient.
A DFT circuit (second L-dimensional mapping means) for mapping to the same L-dimensional distortion evaluation space as that described above, 6 is a changeover switch for switching the DFT sound source vector 2 and input to the sound source vector selection circuit 9, 9 is a changeover switch 6 for selection A source vector selection circuit that selects one optimal source vector code that minimizes the amount of distortion for the DFT input speech 4 using the evaluation weighting filter coefficient 5 and the M mapped source vectors from the M mapped source vectors. (Sound source vector selecting means) and 10 are optimal sound source vector codes selected by the sound source vector selecting circuit 9.

Next, the basic operation of the conventional device will be described. First, the changeover switch 6 transmits the M DFT excitation vectors 2 in the codebook 1 to the optimal excitation vector selection circuit 9 one by one. The optimal sound source vector selection circuit 9 is composed of M DFT sound source vectors 2
For each of them, the distortion amount of the reproduced voice with respect to the input voice in the frequency domain is calculated using the DFT sound source vector 2, the evaluation weight filter coefficient 5, and the DFT input voice 4. The distortion amount D (k) when the k-th sound source vector out of M is used is given by the following equation.

Here, X (i) is the i-th component of the DFT input voice, and H
(I) is the ith component of the evaluation weight filter, C (i, k)
Is the i-th component of the k-th DFT sound source vector, g (k)
Is a gain coefficient that minimizes D (k). Further, according to the above-mentioned document 2, the expression (1) is equivalent to the following expression (2), and the expression (2) is used for the actual calculation.

Here, Y ^* (i) represents a complex conjugate of Y (i),
Y (i) is given by the following equation (3).

 Y (i) = X (i) · a (i) / H (i) (3) a (i) is given by the following equation (4).

a (i) = | H (i) | (4) Give the minimum value among the M D (k) obtained in this way
The number of the DFT sound source vector is selected as the optimum sound source vector code.

[Problems to be solved by the invention]

Since the conventional optimal sound source vector search device is configured as described above,
It is necessary to perform L-dimensional distortion amount calculation M times, and M is increased to obtain a good reproduced sound (for example, M = 1024).
Thus, there is a problem that the amount of calculation required for calculating the amount of distortion is enormous, and the scale of the device when the device is implemented becomes very large.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and an object of the present invention is to provide an optimal sound source vector search device capable of reducing the amount of calculation required for calculating the amount of distortion in the optimum sound source vector search.

[Means for solving the problem]

An optimal excitation vector search apparatus according to the present invention comprises a mapped excitation codebook (1), a speech analysis means (16), a first L-dimensional mapping means (12), and a second L-dimensional mapping means (13). When,
In the optimal excitation vector search device of the speech encoding device including the excitation vector preliminary selection means (3) and the excitation vector selection means (9), the mapped excitation codebook (1) is an L-dimensional orthogonal orthogonal condition. It has M mapped sound source vectors (2) mapped to the distortion evaluation space, and the first L-dimensional mapping means (12) converts the input speech signal (11) into the same L as the mapped sound source vector (2). The sound source vector preliminary selecting means (3) is mapped to the dimensional distortion evaluation space, and is mapped as a mapped audio signal (4).
The voice signal is output to the sound source vector selection means (9), and the voice analysis means (16) analyzes the input voice signal (11) to calculate a synthesis filter coefficient, and calculates an impulse response of the synthesis filter coefficient. To the second L-dimensional mapping means (13), and the second L-dimensional mapping means (13) maps the impulse response of the synthesis filter coefficient to the same L-dimensional distortion evaluation space as the mapped sound source vector (2). Then, it outputs to the sound source vector preliminary selecting means (3) and the sound source vector selecting means (9) as the evaluation weight filter coefficient (5), and the sound source vector preliminary selecting means (3) outputs the evaluation weight filter coefficient (5). L1 based on the magnitude of the absolute value of each component
(L1 <L) dimension components are selected, and the L1 dimension components are subjected to the mapped speech signal (4) using the evaluation weight filter coefficient (5) and the M mapped sound source vectors (2). The M1 (M1 <M) mapped sound source vectors having a small distortion amount are selected, and the sound source vector selecting means (9) selects the sound source vector preliminary selecting means (3) for all L dimensions.
From the M1 mapped sound source vectors, one optimal mapped sound source vector that minimizes the amount of distortion for the mapped speech signal (4) is selected using the evaluation weight filter coefficient (5) and the M1 mapped sound source vectors. It was made.

[Action]

In the present invention, by configuring as described above, the sound source vectors are preliminarily selected, and a mapped sound source vector having a small amount of distortion with respect to the mapped sound signal is selected from among them.
Since the optimum mapped sound source vector that minimizes the amount of distortion for the mapped audio signal is selected from the selected mapped sound source vectors, the amount of calculation required for calculating the amount of distortion when searching for the optimum sound source vector is reduced. .

〔Example〕

 An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing an apparatus for searching for an optimal excitation vector according to an embodiment of the present invention. In FIG. 1, the same reference numerals as those in FIG. 2 denote the same or corresponding parts. Reference numeral 3 denotes a sound source vector preliminary selection circuit (a sound source vector preliminary selection means), which is based on the magnitude of the absolute value of each component of the evaluation weighting filter coefficient 5 from the DFT source vector 2 selected by the changeover switch 6. In addition to selecting L1 (L1 <L) dimensional components, M1 having a small amount of distortion with respect to the DFT input voice 4 is used for the L1 dimensional components by using an evaluation weighting filter coefficient 5 and M DFT excitation vectors 2. (M1 <M) DFT sound source vectors are selected. A changeover switch 6 switches the DFT excitation vector 2 and outputs the DFT excitation vector 2 not to the excitation vector selection circuit 9 but to the excitation vector preselection circuit 3. Reference numeral 8 denotes a second changeover switch, and the sound source vector preliminary selection circuit 3
The DFT excitation vector 2 is switched in accordance with the designation signal 7 output by the.

 Next, the operation will be described.

First, the changeover switch 6 transmits the DFT excitation vector 2 in the codebook 1 to the excitation vector preliminary selection circuit 3. The sound source vector preliminary selection circuit 3 calculates a (i) ·
It is assumed that a dimension having a large | Y (i) | is a dimension having a large contribution to the amount of distortion. First, a large L1 of a (i) · | Y (i) |
Choose dimensions. And M DFT sound source vectors 2
For each of these L1 dimensions, the amount of distortion that the reproduced voice has for the input voice in the frequency domain is calculated using the DFT sound source vector 2, the evaluation weight filter coefficient 5, and the DFT input voice 4. The distortion amount D1 (k) when the k-th sound source vector in M is used is given by the following equation.

Here, I (i) is a vector {| a (j) |, j = 1, L1}.
Is the dimension corresponding to the i-th largest vector component. Of the M D1 (k) obtained in this way, a small D1
The number of M1 sound source vectors giving (k) is sent to the second changeover switch 8 as a sound source vector designating signal 7, and the second changeover switch outputs M1 DFTs giving small D1 (k).
The sound source vectors are transmitted to the sound source vector selection circuit 9 one by one. The following operation of the sound source vector selection circuit 9 is the same as that of the sound source vector selection circuit 9 shown in FIG. 2 except that M1 DFT sound source vectors are selected. I do.

Next, the calculation amount will be described. D (k) or D1 (k)
In the conventional technique, that is, a method of calculating a distortion amount for selecting an optimal sound source vector without performing preliminary selection, the calculation amount of LMF According to the present embodiment, first, L1 ・ M ・ F for the preliminary selection and the sum of L ・ M1 ・ F for the final selection (L1 本 M)
+ L · M1) · F is required, so L1 · M + L
If L1 and M1 are determined so as to satisfy M1 <LM, the amount of calculation can be reduced. At this time, the computation amount decreases as M1 and L1 are smaller, but the optimal excitation vector may not be preselected by the excitation vector preliminary selection circuit.
M1 and L1 need to be determined appropriately. As an experimental example, M = 1
When 024, L = 40 and M1 = 32, L1 = 5, it has been confirmed that the amount of calculation is greatly reduced without the optimal excitation vector being omitted from the preliminary selection result.

In the above-described embodiment, an example in which the process of searching for an optimum sound source is realized in a circuit has been described. However, this may be realized by software processing using a general-purpose arithmetic device such as a microprocessor or a signal processing processor.

In the above embodiment, the case where the frequency domain by DFT is used as the distortion evaluation space has been described. However, any mapping space that satisfies the dimensional orthogonality condition may be used.

〔The invention's effect〕

As described above, according to the present invention, the mapped excitation codebook (1), the speech analysis means (16), the first L-dimensional mapping means (12), and the second L-dimensional mapping means (13) In the optimal excitation vector search apparatus of the speech encoding apparatus comprising the excitation vector preliminary selection means (3) and the excitation vector selection means (9), the mapped excitation codebook (1) satisfies the L-dimensional orthogonal condition. It has M mapped sound source vectors (2) mapped to the L-dimensional distortion evaluation space, and the first L-dimensional mapping means (12) converts the input audio signal (11) into the same as the mapped sound source vector (2). And outputs it to the sound source vector preliminary selecting means (3) and the sound source vector selecting means (9) as a mapped sound signal (4), and the sound analyzing means (16) is inputted. Analyzing the audio signal (11) to calculate the synthesis filter coefficient and The impulse response of the synthesis filter coefficient is calculated and output to the second L-dimensional mapping means (13), and the second L-dimensional mapping means (13) converts the impulse response of the synthesis filter coefficient into the mapped sound source vector (2). Are mapped to the same L-dimensional distortion evaluation space as described above, and output to the excitation vector preselection means (3) and the excitation vector selection means (9) as evaluation weight filter coefficients (5), and the excitation vector preselection means (3) Is L1 based on the magnitude of the absolute value of each component of the evaluation weight filter coefficient (5).
(L1 <L) dimension components are selected, and the L1 dimension components are subjected to the mapped speech signal (4) using the evaluation weight filter coefficient (5) and the M mapped sound source vectors (2). The M1 (M1 <M) mapped sound source vectors having a small distortion amount are selected, and the sound source vector selecting means (9) selects the sound source vector preliminary selecting means (3) for all L dimensions.
From the M1 mapped sound source vectors, one optimal mapped sound source vector that minimizes the amount of distortion for the mapped audio signal (4) is selected using the evaluation weight filter coefficient (5) and the M1 mapped sound source vectors. Therefore, the amount of calculation can be reduced when calculating the amount of distortion for the sound source vector, and even a small-scale device can search for an optimal sound source vector from M sufficiently large sound source vectors. The effect of obtaining higher-quality reproduced audio on a large scale is obtained.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an optimum sound source vector searching device according to an embodiment of the present invention, and FIG. 2 is a block diagram showing a conventional optimum sound source vector searching device. In the figure, 1 is a codebook composed of M DFT excitation vectors, 2 is a DFT excitation vector, 3 is an excitation vector preliminary selection circuit, 4 is a DFT input voice, 5 is an evaluation weight filter coefficient, 6 is a changeover switch, 7 is a designation signal, 8 is a second switch, 9 is a sound source vector selection circuit, and 10 is an optimum sound source vector code. In the drawings, the same reference numerals indicate the same or corresponding parts.

Continuation of front page (56) References JP-A-59-99496 (JP, A) JP-A-62-139089 (JP, A) JP-A-59-77730 (JP, A) JP-A-59-94936 (JP, A) , A)

Claims (1)

(57) [Claims] 1. A mapped excitation codebook, a speech analysis means, and a first
, A second L-dimensional mapping means, an excitation vector preliminary selection means, and an optimal excitation vector search apparatus of a speech coding apparatus including the excitation vector selection means, wherein the mapped excitation codebook has an L-dimensional It has M mapped sound source vectors mapped to an L-dimensional distortion evaluation space that satisfies the three-dimensional orthogonal condition, and the first L-dimensional mapping means converts the input speech signal into the same L-dimensional distortion evaluation space as the mapped sound source vector. The audio signal is mapped and output to the sound source vector preliminary selecting means and the sound source vector selecting means as a mapped sound signal. The sound analyzing means analyzes the input sound signal to calculate a synthesis filter coefficient, and impulse of the synthesis filter coefficient. Calculates the response and outputs it to the second L-dimensional mapping means. The second L-dimensional mapping means converts the impulse response of the synthesis filter coefficient into the same L as the mapped sound source vector. It maps to the dimensional distortion evaluation space and outputs it to the sound source vector preliminary selecting means and the sound source vector selecting means as the evaluation weight filter coefficient. The sound source vector preliminary selecting means determines the magnitude of the absolute value of each component of the evaluation weight filter coefficient. L1 (L1 <L) dimension components are selected based on the L1 dimension components.
Using the evaluation weighting filter coefficient and the M mapped sound source vectors, select M1 (M1 <M) mapped sound source vectors having a small amount of distortion to the mapped sound signal. The sound source vector selecting means targets all L dimensions, From the M1 mapped sound source vectors selected by the sound source vector preliminary selecting means, one optimal mapped sound source vector with the least amount of distortion to the mapped sound signal is selected using the evaluation weighting filter coefficient and the M1 mapped sound source vectors. An optimal sound source vector search device.