JPH0990989A - Conversion encoding method and conversion decoding method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve an efficiency of an entire encoder by encoding a signal with a high efficiency by making use of a regularity of a spikelike pitch component appearing in a frequency domain signal transformed into a frequency domain. SOLUTION: An encoder A is constituted of a time-frequency converter 1, a general situation outline calculation and quantization device 2, a first flattening device 3, a pitch encoder 4 an adder 5, a minute spectrum outline calculation and quantization device 6, a second flattening device 7, and a quantization device 8. And a decoder B is constituted of a reproducing device 9, a minute spectrum outline reproducing device 10, a first inverse flattening device 11, a pitch reproducing device 12, an adder 13, a general situation outline reproducing device 14, a second inverse flattening device 15, and a time-frequency inverter 16. And a tone signal or a voice signal is divided into frames of a certain interval and is converted into a frequency domain signal. Next, only the pitch component is extracted from the frequency domain signal, to be separated and decoded.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a transform coding method and a transform decoding method for a signal including a pitch component such as a tone signal or a voice signal.

[0002]

2. Description of the Related Art At present, as a method of highly efficiently encoding an audio signal such as a musical tone signal or a voice signal, the audio signal is divided into intervals of a constant interval of about 5 to 50 ms called a frame, Time to signal-
The frequency domain signal obtained by performing frequency conversion is the envelope shape of the frequency characteristic (frequency characteristic outline) and the residual signal obtained by flattening the frequency domain signal with the frequency characteristic outline.
It has been proposed to separate the information into two pieces and encode each piece.

As a concrete method of such an encoding method, an adaptive spectral perceptual control entropy encoding method (ASPEC, Adaptive Spectral Perceptual Entropy Codin) is used.
g), transform coding method by weighted vector quantization (TCW
VQ, Transform Coding with Weighted Vector Quantizati
on), and Mpeg-Audio Layer 3 method (MP
EG-Audio Layer 3) etc. have been proposed.

Regarding these technologies, K. Brande
nburg, J.Herre, JDJohnston et al: "ASPEC: Adaptive
spectral entropy coding of high quality music sign
als ", Proc.AES'91, T.Moriya, H.Suda:" An 8 Kbit / s
transform coder for noisychannels ", Proc.ICASSP'89
pp.196--199, and ISO / IEC standard IS-11172-3.

Here, in order to realize highly efficient coding by these coding methods, it is desirable that the residual signal has as flat a frequency characteristic as possible. For this reason, the adaptive spectrum auditory control entropy coding method (AS
PEC) or Mpeg-Audio Layer 3 method (M
In PEG-Audio Layer 3), as shown in FIG. 5, the frequency domain signal is divided into several small bands, and the signal in each small band is divided by a value called a scaling factor that represents the strength of the band. By normalizing, the frequency characteristic of the residual signal is flattened.

On the other hand, as a method of flattening a frequency domain signal which is more efficient than these methods, there is a method using linear prediction analysis as shown in FIG. In this method, the frequency characteristics are flattened by driving a linear prediction analysis filter with a linear prediction coefficient obtained by linearly predicting an input signal.
This method is a method used in the transform coding method (TCWVQ) by the weighted vector quantization.

Regarding each related technique such as linear prediction analysis, discrete cosine transform (DCT) and modified discrete cosine transform (MDCT), Saito and Nakata "Basics of Speech Information Processing"
(Ohmsha) Chapter 6, KRRao, P.Yip, Translated by Yasuda and Fujiwara "Image Coding Technology-DCT and Its International Standards" Chapter 2 (Ohmsha), HSMalvar, "Signal Processing with L
apedTransforms, "Artech House, and ISO / IEC standard IS-11172-3.

[0008]

However, in these encoding methods, the general outline of the frequency characteristic is limited to normalization, and the unevenness of the microscopic frequency characteristic due to the pitch component of the musical sound or voice is efficiently generated. It cannot be removed well. Therefore, this becomes an obstacle, and it is difficult for the above-mentioned conventional encoding method to improve the efficiency when encoding an audio signal having a strong pitch component.

The present invention has been made in view of the above-mentioned problems, and provides a transform coding method and a transform decoding method capable of efficiently coding an audio signal containing a pitch component. It is an object.

[0010]

According to the first aspect of the present invention,
A conversion coding method of a tone signal or a voice signal including a pitch component, which divides a voice signal or a voice signal into frames at constant time intervals and transforms into a frequency domain signal.
, A second step of extracting only a pitch component from the frequency domain signal, separating and encoding the same, and a third step of encoding a signal from which the pitch component is removed from the frequency domain signal. Is characterized by.

According to a second aspect of the present invention, in the first aspect of the invention, the second step is the fourth step of obtaining and encoding the fundamental frequency of the pitch component, and the pitch based on the fundamental frequency. And a fifth step of extracting and encoding the component.

According to a third aspect of the invention, in the second aspect of the invention, the fifth step is such that a sample of a frequency domain signal closest to a frequency that is a natural number multiple of the fundamental frequency is centered, and is included continuously. It is characterized in that the pitch component is extracted with a plurality of samples as a unit.

The invention according to claim 4 is the invention according to claim 2 or 3.
In the invention described above, the fifth step is characterized in that each unit of pitch components is encoded by vector quantization.

The invention described in claim 5 is a method for decoding a code obtained by the transform coding method according to claim 1, wherein the pitch component coded in the second stage is decoded. Step, a seventh step of decoding a signal obtained by removing a pitch component from the frequency domain signal encoded in the third step, a composite output obtained in the sixth step and a seventh step The eighth step of transforming the frequency domain signal obtained by synthesizing the composite output obtained in 1) into a time domain signal.

According to a sixth aspect of the invention, in the fifth aspect of the invention, the sixth step includes a ninth step of decoding the fundamental frequency of the pitch component obtained in the fourth step, and the ninth step. 9
Arranging the pitch component as a signal in the frequency domain based on the fundamental frequency of the pitch component obtained by the step of
It is characterized in that

According to a seventh aspect of the invention, in the tenth aspect of the invention, the tenth step is such that a continuous sample including a sample in a frequency region closest to a natural multiple of the fundamental frequency is taken as one unit. The feature is that pitch components are arranged.

The invention according to claim 8 is the invention according to claim 6 or 7.
In the described invention, the tenth step is characterized in that the vector-quantized index is decoded for each unit of the pitch component.

[0018]

The musical tone or voice has a pitch, that is, high / low pitch. The frequency domain signal obtained by frequency-converting this musical sound or voice contains pitch components arranged at regular frequency intervals. Therefore, the above-mentioned pitch component is also included in the residual signal obtained by normalizing the frequency domain signal with the outline of its own frequency characteristic. Since this pitch component appears as a spike having a large energy with respect to the total power, it lowers the flatness of the residual signal and deteriorates the quantization efficiency. However, the present invention focuses on the fact that the pitch components are arranged at equal intervals on the frequency axis, and subtracts the pitch components from the residual signal, thereby increasing the flatness of the residual coefficient with a small amount of additional information.

[0019]

DETAILED DESCRIPTION OF THE INVENTION An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram for explaining the transform coding method and transform decoding method according to the present embodiment, where code A is an encoder and B is a decoder. As shown in the figure, the encoder A includes a time-frequency converter 1, a global rough shape calculator / quantizer 2, a first flattener 3, a pitch encoder 4, an adder 5, and a fine spectrum rough shape calculation.・ Quantizer 6,
The second flattener 7 and the quantizer 8 are included.

The time-frequency converter 1 is an input signal in the time domain (audio signal such as tone signal or voice signal).
Is divided into frames having a constant time interval, and each frame is subjected to time-frequency conversion to generate a frequency domain signal.

FIG. 2 shows the frequency characteristic of this frequency domain signal. As shown in this figure, the frequency domain signal of the tone signal or the voice signal has a constant frequency interval p.
It contains the pitch components arranged in. The conversion method is Discrete Cosine Transf
ormation, DCT) and modified discrete cosine transform (Modified Dis)
crete Cosine Transformation (MDCT) can be used.

The global outline calculator / quantizer 2 generates and quantizes a signal representing the global outline of the frequency domain signal output from the time-frequency converter 1. Then, while outputting this signal to the first flattener 3,
Output as a quantized global outline index.
As a method of calculating the global outline, a linear prediction spectrum or a scale factor that divides the frequency domain signal into a plurality of subbands and expresses the overall shape of the frequency domain signal by the representative value of each band may be used. Good.

When the linear prediction spectrum is quantized, the linear prediction parameters are converted into LSP parameters and quantized. Alternatively, it is converted into K parameters and quantized.

The first flattener 3 calculates a global shape of the frequency domain signal output from the time-frequency converter 1
It is flattened by division by the global outline signal output from the quantizer 2, and a first flattened signal is output.

Next, the pitch encoder 4 detects the pitch component from the first flattened signal and encodes it. Also,
FIG. 3 is a diagram showing the details of the pitch encoder 4, in which the first flattened signal is the pitch fundamental frequency extractor 4 shown.
a and the pitch sample extractor 4b.

This pitch fundamental frequency extractor 4a has a first
The fundamental frequency of the pitch component (pitch fundamental frequency) is obtained by analyzing the flattened signal of. That is, the pitch fundamental frequency extractor 4a calculates the cepstrum of the first flattening coefficient, and sets the maximum value as the fundamental cycle of the pitch component. Then, the pitch fundamental frequency is calculated by calculating the reciprocal of the fundamental period, and the pitch fundamental frequency quantizer 4c
Output to

In order to make the pitch fundamental frequency more accurate, the fundamental frequency at which the total sum of the powers of the samples of the first flattening signal for each pitch fundamental frequency is maximized before and after the obtained pitch fundamental frequency. You may search and use this as a new pitch fundamental frequency.

The pitch fundamental frequency quantizer 4c quantizes the pitch fundamental frequency thus obtained. That is, the pitch fundamental frequency quantizer 4c scalar-quantizes the logarithmic value of the pitch fundamental frequency and outputs it to the outside as a quantized pitch fundamental frequency index, and at the same time, outputs the scalar-quantized signal to the pitch sample extractor 4b. Output to.

The pitch sample extractor 4b is a natural number of the quantized pitch fundamental frequency input from the pitch fundamental frequency quantizer 4c with respect to the first flattened signal input from the first flattener 3. One sample before and after is extracted centering on the sample closest to the doubled frequency, and a set of these three samples is output to the pitch sample quantizer 4d as a sample group of one pitch component. The number of sample groups of this pitch component may be a fixed value or may be variable.

The pitch sample quantizer 4d quantizes the pitch component sample group and outputs it to the outside as a quantized pitch component index, and the quantized pitch component obtained by decoding the quantized pitch component index is added to the adder. Output to 5. The quantization of the sample group is
Scalar quantization may be used, or vector quantization may be performed for each sample group consisting of three samples. Further, vector quantization may be performed on all sample groups at once. The above is the processing performed in the pitch encoder 4.

Next, the adder 5 uses the quantized signal of the pitch component input from the pitch encoder 4 to generate the pitch component from the first flattened signal input from the first flattener 3. A second flattening signal is generated by subtracting only the above, and is output to the fine spectrum rough shape calculating / quantizing unit 6 and the second flattening unit 7.

FIG. 4 is a diagram showing the frequency characteristic of the second flattened signal. As can be seen from the comparison with FIG. 2, the second flattened signal is the frequency domain signal output from the time-frequency converter 1 with the pitch component removed.

The fine spectrum rough shape calculator / quantizer 6 is
A fine spectrum outline (fine spectrum outline) is calculated from the second flattened signal and quantized. Then, this quantized signal is output to the outside as a quantized fine spectrum outline index, and is also output to the second flattener 7.

The fine spectrum outline may be obtained by directly quantizing the fine spectrum outline, or may be obtained by linearly combining the fine spectrum outlines of past frames.
Alternatively, the information of the quantized fine spectrum outlines of the past and present frames may be obtained by linear synthesis. Further, as this fine spectrum outline, for example, a value obtained by convolving a window function having a width of about 3 to 5 with the absolute value of the second flattened signal may be used, or the second flattened signal obtained by subband division. It is also possible to prepare a representative value of the amplitude of the encoded signal for each band and use this as a rough shape.

The second flattening unit 7 divides the second flattening signal input from the adder 5 by the fine spectrum rough shape obtained by the fine spectrum rough shape calculating / quantizing unit 6 to flatten it,
The third flattened signal is output to the quantizer 8. The quantizer 8 performs scalar quantization or vector quantization on the third flattened signal and outputs it to the outside as a quantization index.

In the case of vector quantization, all samples of a frame may be quantized at once, but it is better to divide the sample sequence of the frame into a plurality of subvectors and quantize each subvector. It is realistic in terms of calculation amount. The division method may be simple subband division, or interleave division in which samples are interleaved and then divided. Also, adaptive bit allocation may be performed according to the amount of information required for quantization.

Next, the decoder B will be described. As shown in FIG. 1, the decoder B includes a regenerator 9, a fine spectrum outline regenerator 10, a first inverse flatter 11, and a pitch regenerator 1.
2, an adder 13, a global rough regenerator 14, a second inverse flatter 15, and a time-frequency inverse converter 16.

Of these, the regenerator 9 regenerates the third flattened signal from the quantization index transmitted from the encoder A. The regenerator 9 reproduces the third flattened signal by performing the inverse process of the quantizer 8 and outputs it to the first inverse flattener 11. The fine spectrum outline regenerator 10 reproduces the fine spectrum outline from the fine spectrum outline quantization index transmitted from the encoder A.

The first inverse flattener 11 adds a fine spectrum outline to the third flattened signal input from the regenerator 9, reproduces the second flattened signal, and supplies it to the adder 13. Output. The pitch regenerator 12 regenerates the pitch component from the quantized pitch component index and the quantized pitch fundamental frequency index transmitted from the encoder A, and outputs it to the adder 13.

The adder 13 adds the pitch component input from the pitch regenerator 12 to the second flattened signal input from the first inverse flattener 11 to reproduce the first flattened signal. , To the second inverse flatter 15. Further, the global outline regenerator 14 reproduces the global outline from the quantized global outline index transmitted from the encoder A, and outputs it to the second inverse flatter 15.

The second inverse flatter 15 adds the global outline input from the global outline regenerator 14 to the first flattened signal input from the adder 13, and outputs the frequency domain signal. To generate. Then, the time-frequency inverse converter 16 performs time-frequency inverse conversion on the frequency domain signal input from the second inverse flatter 15 and decodes it to output a time domain voice signal or musical tone signal.

[0042]

As described above, according to the present invention,
When encoding a musical tone signal or a voice signal having a pitch component, the regularity of spike-like pitch components appearing in a frequency domain signal obtained by converting the signal into the frequency domain is used to highly efficiently encode the signal. . Therefore, a more flattened residual coefficient can be obtained, and the efficiency of the entire encoder can be improved.

[Brief description of drawings]

FIG. 1 is a diagram illustrating an encoder and a decoder according to an embodiment of the present invention.

FIG. 2 is a diagram showing frequency characteristics of an output signal of the time-frequency converter in the present invention.

FIG. 3 is a diagram showing a detailed configuration of a pitch encoder in the present invention.

FIG. 4 is a diagram showing frequency characteristics of a second flattened signal in the present invention.

FIG. 5 is a first diagram illustrating a conventional transform coding method.

FIG. 6 is a second diagram illustrating a conventional transform encoding method.

[Explanation of symbols]

 1 Time-Frequency Converter 2 Global Approximate Calculator / Quantizer 3 First Flatter 4 Pitch Encoder 5 13 Adder 6 Fine Spectrum Approximate Calculator / Quantizer 7 2nd Flatter 8 Quantum 9 Regenerator 10 Fine spectrum rough regenerator 11 First inverse flatter 12 Pitch regenerator 14 Global rough regenerator 15 Second inverse flatter 16 Time-frequency inverse converter

Claims (8)

[Claims] 1. A method for transform-encoding a tone signal or voice signal containing a pitch component, the method comprising: a first step of dividing the tone signal or voice signal into frames at fixed time intervals and transforming into a frequency domain signal; It is characterized by comprising a second step of extracting only the pitch component from the frequency domain signal, separating and encoding the same, and a third step of encoding a signal from which the pitch component is removed from the frequency domain signal. Transform coding method. 2. The second step comprises a fourth step of obtaining and coding a fundamental frequency of a pitch component and a fifth step of extracting and coding a pitch component based on the fundamental frequency. The transform coding method according to claim 1, wherein 3. A fifth step is to extract a pitch component with a unit of a plurality of continuous samples including a sample of a frequency domain signal closest to a frequency which is a natural multiple of the fundamental frequency. The transform coding method according to claim 2. 4. The transform coding method according to claim 2, wherein in the fifth step, coding is performed by vector quantization for each unit of the pitch component. 5. A method for decoding a code obtained by the transform coding method according to claim 1, comprising a sixth step of decoding the pitch component coded in the second step, and a third step. A seventh step of decoding a signal from which a pitch component has been removed from the frequency domain signal encoded in the step of :, the composite output obtained in the sixth step, and the composite output obtained in the seventh step. An eighth step of transforming a frequency domain signal obtained by synthesizing outputs into a time domain signal, the transform decoding method. 6. The sixth step also includes a ninth step of decoding the fundamental frequency of the pitch component obtained in the fourth step and a fundamental frequency of the pitch component obtained by the ninth step. First, the pitch component is arranged as a signal in the frequency domain in and
6. The transform decoding method according to claim 5, further comprising: 7. The pitch component is arranged in the tenth step with one continuous sample including a sample in a frequency region closest to a frequency that is a natural multiple of the fundamental frequency as one unit. Conversion decoding method. 8. The transform decoding method according to claim 6, wherein in the tenth step, the vector-quantized index is decoded for each unit of the pitch component.