JPH0669809A - Voice coding device - Google Patents

Abstract

PURPOSE:To provide the MBE system voice coding device which enables coding at a low bit rate. CONSTITUTION:This device is provided with a first compensator 1 to compensate fixed-dimensional data from the input data of a spectrum amplitude value, vector quantizer 4 to perform vector quantization based on the compensated fixed-dimensional data, corrector 7 to correct quantization error with the vector quantization, and scolar quantizer 8 to perform scalar quantization to the correcting amount of this correction. The vector quantization of data is enabled by fixing the dimension of data, and a system for calculating difference from a predictive value by utilizing correlation between frames can be applied. The quantization error can be completely corrected by correcting one representative point in each block.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice coder used in a digital telephone or a digital recorder which requires high-efficiency voice compression, and more particularly to a voice coder capable of low bit rate voice coding. .

[0002]

2. Description of the Related Art In audio coding, it is required to code audio signals with as few bits as possible in order to effectively use frequencies and storage media.

A voice encoding device using the MBE (Multi-Band Excitation) system is considered promising as a device capable of performing voice encoding at a low bit rate. M
The BE system divides an input speech into a plurality of bands according to a fundamental frequency, determines voiced / unvoiced for each band, and quantizes a spectrum amplitude. A speech coding apparatus using this MBE system is As shown in FIG. 7, a voice analysis unit 10 that analyzes an input voice, a fundamental frequency quantization unit 20 that outputs quantized data relating to a fundamental frequency, a voiced or unvoiced band is determined, and the result is quantized. Voiced / unvoiced decision quantizing unit 30 for converting the data into quantized data, a spectrum amplitude quantizing unit 40 for quantizing the spectrum amplitude of each band, and a multiplexing unit 50 for multiplexing the output of each quantizing unit. Therefore, the input voice is coded with about 20 ms, which is considered to be almost constant, as one frame.

Of these, the spectral amplitude quantizer 40 requires the most number of bits for quantization.

As shown in FIG. 8, a spectrum amplitude quantizing unit 40 of a conventional MBE system speech coding apparatus includes a spectrum amplitude band divider 41 for dividing an input spectrum into a plurality of (n) bands, and a signal. Discrete cosine transform (DCT) device 42 for orthogonal transforming, a coefficient discriminator 43 for extracting a direct current (DC) component, which is a coefficient representing an average of spectral amplitude values for each band, from the DCT coefficient, and each coefficient discriminator. A vector synthesizer 44 that synthesizes one n-order vector having n extracted components of 43 as an element, a vector quantizer 45 that quantizes the vector output from the vector synthesizer 44, and an error of each DCT coefficient And a scalar quantizer 46 for quantizing.

The spectrum amplitude input to the spectrum amplitude quantizer is first divided into n bands by the spectrum amplitude divider 41, and the divided spectrum amplitudes are discrete cosine transform ( DCT).
The DCT coefficient obtained here represents frequency spectrum information. This DCT coefficient is discriminated in the coefficient discriminator 43 into a coefficient representing the average of the spectrum amplitude value for each band and a coefficient other than that. The former is synthesized as an n-dimensional vector by the vector synthesizer 44 and then vector-quantized by the vector quantizer 45. On the other hand, the latter is scalar-quantized by the scalar quantizer 46 for each coefficient and output as a value representing an error.

[0007]

However, the conventional MBE
The system speech coding apparatus applies vector quantization, which is an efficient low bit rate quantization method, to the DC component of each band, but does not directly vector quantize each coefficient. It is not. This is because, in the MBE system speech encoding device, the number of data representing the spectrum amplitude value changes for each frame depending on the fundamental frequency of the inputted speech, so that the dimension of the vector is not fixed and the vector quantization is directly applied. Because you can't. Therefore,
It can be considered that the conventional MBE-type speech encoding apparatus basically performs scalar quantization, and there is room for improvement in order to improve the efficiency of quantization.

The present invention solves the above-mentioned conventional problems, and by constructing the vector quantization so that it can be directly applied, speech of the MBE system which enables coding at a low bit rate. An object is to provide an encoding device.

[0009]

Therefore, in the present invention, M
In a BE system speech encoding device, an interpolation means for interpolating fixed-dimensional data from input data of a spectrum amplitude value, a vector quantization means for performing vector quantization based on the interpolated fixed-dimensional data, and a vector quantization There is provided a correction means for correcting the quantization error associated with the above, and a scalar quantization means for performing scalar quantization on the correction amount in this correction.

Further, in order to create the data to be supplied to the vector quantizing means, the predicting means for predicting the fixed dimension data of the current frame on the basis of the fixed dimension data of the previous frame, and the predictive data and the output of the interpolation means. A difference means for calculating the difference from the data is provided, and the output data of the difference means is supplied to the vector quantization means.

Further, in order to create data to be supplied to the correction means, an addition means for adding the vector quantized vector element and the prediction data output by the prediction means, and a spectrum based on the data output by the addition means A second interpolating means for decoding the amplitude value is provided, and in the correcting means,
The correction is performed so that the error between the output data of the second interpolation means and the input data of the spectrum amplitude value is minimized.

[0012]

In this speech coding apparatus, fixed-dimensionalized spectrum amplitude value data is created by interpolation from input data of the spectrum amplitude value. By making the data fixed dimensions in this way, vector quantization of the data becomes possible.

Further, the fixed dimension makes it possible to utilize the correlation between frames, and a method for obtaining the difference from the predicted value can be applied.

Further, in correcting the quantization error,
The spectrum amplitude value is decoded from the quantized data, and correction is performed so that the error between the decoded value and the actual input data of the spectrum amplitude value is minimized. All you have to do is correct one representative point.

[0015]

BEST MODE FOR CARRYING OUT THE INVENTION As shown in FIG. 1, a MBE-type speech encoding apparatus according to an embodiment of the present invention interpolates a fixed-dimensional spectrum amplitude value from an input spectrum amplitude value as shown in FIG. The interpolator 1, the predictor 2 that predicts the spectrum amplitude value of the current frame from the spectrum amplitude value of the previous frame, the differencer 3 that obtains the error between the interpolated amplitude value and the predicted value, and the output difference as elements. And a vector quantizer 4 for quantizing the vector, and as a configuration for encoding the error amount associated with the vector quantization, the quantized vector and the predicted value are added to obtain a fixed-dimensional spectrum amplitude value. , A second interpolator 6 for decoding the spectrum amplitude from the calculated spectrum amplitude value, the decoded spectrum amplitude and the actually input spectrum amplitude value Corrector 7 error between the corrected one of the spectral amplitude value of the fixed dimension so as to minimize
And a scalar quantizer 8 for scalar-quantizing the correction amount obtained by the corrector 7.

The operation of the spectrum amplitude quantizer will be described with reference to FIGS.

The data at the N ₁ point of the inputted spectrum amplitude value is indicated by a black circle in FIG. In the first interpolator 1,
Based on the data of N ₁ points, the spectrum amplitude value (white circle) of N ₂ points passing through given data points (points on the dotted line) is interpolated. The spectrum amplitude value of this white circle becomes fixed-dimensional data.

The predictor 2 predicts the amplitude value in the next frame from the obtained fixed-dimensional spectrum amplitude value.

On the other hand, the differencer 3 obtains the difference between each spectrum amplitude value in the fixed dimension and the predicted value obtained from the previous frame, and the vector quantizer 4 obtains the vector whose elements are the difference in the fixed dimension. Quantize.

The adder 5 adds the element of the quantized vector and the predicted amplitude value to calculate the spectrum amplitude value in the fixed dimension. The spectrum amplitude value obtained by this quantization includes a quantization error associated with vector quantization, as indicated by Δ in FIG.

In order to create this quantization error correction data, first, the original spectrum amplitude value is decoded using the spectrum amplitude value data output from the adder 5 and obtained by quantization. Therefore, as shown in FIG.
A plurality of interpolation sections including some data are set, and the second interpolator 6 is used to perform interpolation for filling the values between the data. The interpolation curve obtained by this interpolation is shown in FIG. In FIG. 5, black circles represent input spectrum amplitude values, and white circles represent spectral amplitude value data obtained by quantization.

Next, as shown in FIG. 6, in each interpolation section, quantization is performed so that the error between the spectrum amplitude obtained by interpolation and the input spectrum amplitude (black circle) is minimized. One of the spectrum amplitude values (white circles) is corrected using the corrector 7.

In FIG. 6, the corrected spectrum amplitude value is indicated by a square. As is clear from the figure, the interpolation curve can be brought close to the input data by only correcting one of the points obtained by the quantization. Since all of the points obtained by the quantization are fixed dimensions and the position in the horizontal axis direction in FIG. 6 is fixed, as the correction information of the quantization error, only the correction amount of one corrected point is notified. Is enough.

The correction amount of the correction performed by the corrector 7 is
The scalar quantizer 8 performs scalar quantization and outputs.

As described above, in the apparatus of the embodiment, since the spectrum amplitude is fixed, the spectrum amplitude value can be directly vector-quantized. Further, it becomes possible to use the correlation between frames, and it is possible to improve the coding efficiency by using the difference. Furthermore, when correcting the quantization error, it is sufficient to correct only one point in the section. All of these contribute greatly to lowering the bit rate in speech coding.

[0026]

As is apparent from the above description of the embodiments, the speech coding apparatus of the present invention can perform speech coding at a low bit rate. Moreover, since the quantization error can be appropriately corrected, highly accurate speech coding can be performed.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of a speech encoding apparatus of the present invention,

FIG. 2 is a diagram for explaining an interpolation operation of a first interpolator in the apparatus of the embodiment,

FIG. 3 is a diagram for explaining a quantization error occurring in the apparatus of the embodiment,

FIG. 4 is a diagram for explaining an operation of dividing an interpolation section in the apparatus of the embodiment,

FIG. 5 is a diagram for explaining an interpolation operation of a second interpolator in the apparatus of the embodiment,

FIG. 6 is a diagram for explaining a correction operation of a corrector in the apparatus of the embodiment,

FIG. 7 is a block diagram showing the configuration of a conventional MBE system speech encoding device,

FIG. 8 is a block diagram showing a configuration of a spectrum amplitude quantization unit of the conventional device.

[Explanation of symbols]

 1 first interpolator 2 predictor 3 differencer 4,45 vector quantizer 5 adder 6 second interpolator 7 corrector 8,46 scalar quantizer 10 speech analysis unit 20 fundamental frequency quantization unit 30 voiced / unvoiced Decision quantizer 40 Spectral amplitude quantizer 41 Spectral amplitude band splitter 42 Discrete cosine transformer 43 Coefficient discriminator 44 Vector combiner 50 Multiplexer

Claims (3)

[Claims] 1. An MBE-type speech encoding apparatus, comprising: interpolating means for interpolating fixed dimension data from spectral amplitude value input data; and vector quantizing means for performing vector quantization based on the fixed dimension data. Correction means for correcting the quantization error associated with the vector quantization,
A speech coding apparatus, comprising: a scalar quantization unit that scalar-quantizes a correction amount in the correction. 2. Predicting means for predicting fixed-dimensional data of a current frame based on fixed-dimensional data of a previous frame, and predicting the predicted data and the interpolation in order to create data to be supplied to the vector quantizing means. The speech coding apparatus according to claim 1, further comprising: a difference means for calculating a difference from the data output by the means, and supplying the output data of the difference means to the vector quantization means. 3. Addition means for adding the elements of the vector-quantized vector and prediction data output by the prediction means to create data to be supplied to the correction means, and data output by the addition means. A second interpolating means for decoding the spectral amplitude value based on the above, and the correcting means so that the error between the output data of the second interpolating means and the input data of the spectral amplitude value is minimized. The speech coding apparatus according to claim 1 or 2, wherein