JP2847730B2 - Audio coding method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a speech coding method, and more particularly to a speech coding method which converts a speech signal to a low bit rate, particularly 4.8 kb / s or less, and achieves high quality with a relatively small amount of computation. The present invention relates to a speech encoding method for encoding.

[Conventional technology]

As a method of encoding an audio signal at a low bit rate of about 4.8 kb / s, for example, Japanese Patent Application No. 59-272435 (Reference 1) and Japanese Patent Application No. 60-178911 (Reference 2) The pitch interpolation multipulse method described is known. In this method, the transmitting side extracts, from the audio signal for each frame, a spectrum parameter representing the spectrum characteristic of the audio signal and a pitch parameter representing the pitch, classifies the audio signal into two types, a voiced section and an unvoiced section. The information is transmitted between the voiced section and the unvoiced section as follows.

That is, in the voiced section, the sound source signal of one frame is
One pitch section (representative section) of a plurality of pitch sections obtained by dividing one frame for each pitch section is represented by a multipulse, and the amplitude, position, spectrum, pitch parameter, and the like of the multipulse in the representative section are represented. Is transmitted. In the unvoiced section, the sound source of one frame is represented by a small number of multi-pulses and a noise signal, and the amplitude, position, gain, and index of the multi-pulse are transmitted.

On the other hand, on the receiving side, in the voiced section, the amplitude and position of the multi-pulses are interpolated using the multi-pulse of the representative section of the current frame and the multi-pulse of the representative section of the adjacent frame, and the pitch section of the pitch section other than the representative section is used. The multipulse is restored, and the driving sound source signal of the frame is restored. In the unvoiced section, the driving sound source signal of the frame is restored using the multipulse, the index of the noise signal, and the gain.

On the receiving side, the restored driving sound source signal is further input to a synthesis filter using spectral parameters to output a synthesized audio signal, and thus the audio signal is reproduced from the receiving side.

[Problems to be solved by the invention]

However, when the audio signal is encoded and the information for reproduction on the receiving side is transmitted as described above, a fixed amount of information is conventionally assigned to each frame, and thus the transmission information amount is fixed. In other words, if the method of allocating a fixed amount of information for each frame is to be transmitted, if it is desired to further reduce the bit rate, the bit rate must be reduced while ensuring the required reproduction quality. Is not easy.

Specifically, in the above-described conventional method, a fixed amount of information (for example, a bit rate of 4.8 kb / s
In the case of, a 96-bit information amount per frame) is allocated. Assigning a fixed amount of information for each frame in this way is convenient for realizing a device, but has the problem that it is difficult to further reduce the bit rate.

SUMMARY OF THE INVENTION An object of the present invention is to solve the above-mentioned problems and to provide a speech encoding method that enables good reproduction of sound quality at a low bit rate with a relatively small amount of calculation.

[Means for solving the problem]

The speech encoding method according to the present invention obtains a parameter representing a feature of the speech signal from an input discrete speech signal, segments the speech signal according to the feature parameter, and a section obtained by the segmentation is a vowel. When other than the section, the order of the spectrum parameter representing the spectrum envelope of the speech signal of the section and the number of multipulses representing the sound source signal of the speech signal are determined, and the spectrum parameter and the amplitude and position of the multipulse are determined. When the section obtained by the segmentation is a vowel section, a coefficient for predicting the signal of the vowel section is obtained from a signal of at least one section adjacent to the vowel section.

[Action]

In the present invention, it has flexibility compared to the method of allocating and transmitting a fixed amount of information for each frame, and the amount of information necessary for human phoneme perception is segmented according to a feature parameter representing a feature of a speech signal. It is possible to allocate and transmit according to the segment. The amount of transmitted information is reduced as compared with conventional ones, and the correlation of waveforms can also be used, so that the bit rate can be further reduced compared to the conventional method, and the perception of phoneme and naturalness can be reduced. It is possible to maintain good sound quality even in a portion where the characteristic of important sound changes.

〔Example〕

FIG. 1 is a block diagram showing one embodiment of a speech encoding method according to the present invention. As shown in FIG. 1, in the present embodiment, an audio encoding device for audio encoding processing is used, which is composed of circuit units below a buffer memory 110 connected to an input terminal 100 to which an audio signal is supplied.

An output terminal 300 is connected to the multiplexer 260 constituting the final stage of the speech coding apparatus having the configuration shown in FIG.
Is transmitted as transmission information. This speech coding apparatus may basically perform coding processing using a multi-pulse coding method, and the multiplexer 260 stores information about the spectral parameters and the amplitude and position of the multi-pulse. Input and these are transmitted.

As described above, the multi-pulse encoding method is basically used. In the configuration shown in FIG.
And a segmentation circuit 120. The feature extraction circuit 115 and the segmentation circuit 120 are used to obtain a characteristic parameter representing the characteristic of the audio signal from the input discrete audio signal, and to segment the audio signal according to the characteristic parameter. .

Further, in addition to the above, the speech coding apparatus includes switches 195 and 235 connected to a line l from the segmentation circuit 120 to the multiplexer 260, respectively. In the present embodiment, information on the segmentation when the audio signal is segmented as described above is transmitted on the line l, and the switches 195 and 23 are output.
5 can be switched synchronously accordingly. That is, the switches 195 and 235 are selectively switched to the terminals 195a and 235a or to the terminals 195b and 235b, respectively.

These switches 195 and 235 determine the order of spectral parameters representing the spectral envelope of the audio signal of the section and the number of multipulses representing the sound source signal of the audio signal when the section obtained by the segmentation is other than the vowel section. A coefficient for predicting a signal of the vowel section from a signal of at least one section adjacent to the vowel section when a section obtained by the segmentation is a spectrum parameter and an amplitude and a position of the multipulse. Used to ask for

In the speech encoding method according to the present invention, when encoding a speech signal and transmitting information, as described above, a parameter representing a feature of the speech signal is obtained from the input discrete speech signal, and the speech is determined according to the feature parameter. Along with segmenting the signal, when the section obtained by the segmentation is other than the vowel section, the order of the spectrum parameter representing the spectrum envelope of the speech signal of the section and the number of multipulses representing the sound source signal of the speech signal are determined. The spectrum parameter and the amplitude and position of the multi-pulse are obtained, and when the section obtained by the segmentation is a vowel section, the signal of the vowel section is obtained from the signal of at least one section adjacent to the vowel section. A coefficient for prediction is obtained.

This is based on the following findings. That is, the speech waveform carries phoneme information, but the amount of information necessary for perception of the phoneme information varies depending on the section of the speech waveform and is not constant. For example, the amount of information necessary for the perception of phonological information differs depending on the vowel stationary part, the transient part, etc., and the information of the vowel stationary part is not so important for the perception of vowels and consonants, and the signal of the vowel stationary part is 0. Vowel intelligibility
100% is obtained, and consonant intelligibility does not decrease. On the other hand, if the information amount in the transient part is reduced to a certain value or less, both the vowel and consonant intelligibility are reduced.

Therefore, the voice is divided into various voice sections (for example, a vowel stationary part,
(Transient part, consonant part, etc.), and by allocating the amount of information necessary for phonological perception according to the characteristics of these sections and transmitting the information, the bit rate can be further reduced as compared with the conventional method.

That is, it is possible to realize a speech encoding method with a bit rate of 4.87 kb / s or less and high sound quality with a relatively small amount of calculation.

To put it in more detail, the principle can be explained as follows.

First, a segmentation circuit as shown in FIG.
When segmentation is performed at 120, for example, the audio signal is roughly segmented (divided) into vowel parts, vowel stationary parts, transient parts, consonant parts, and silent parts. This includes
For example, it is possible to use characteristic parameters such as the power, spectrum change rate, and pitch of the audio signal obtained for each short section of about 5 msec, and to further use a temporal change of these parameters in a section of 20 to 30 msec. .

A specific segmentation is, for example, a vowel section when the power is large, the spectrum change rate is small, the temporal change between the power and the spectrum change rate is small, and the pitch is clear, and further, a section of a certain length of time in the center of the vowel section. The vowel stationary part, and the segment before and after the maximum point of the spectral change rate where the temporal change of the power and the spectral change rate is large is the transient part, and the power and the temporal change of the power and the spectral change rate are further small. Segment the consonants when not too big, and silence when the power is very low. In this case, the segmentation accuracy is 5 msec, but this can be set to any value by changing the calculation section length of the feature parameter.

As the segmentation method, a well-known method, for example, used in the field of speech recognition can be used in addition to such a method. The method of obtaining the above-mentioned spectral change rate is described in a paper entitled "The role of spectral change information in speech perception" by Furui (Acoustic Society of Japan, H-85-6, 1985).
(Reference 3) can be referred to.

The segmented sections are encoded as follows based on the multi-pulse encoding method.

That is, when the segmented section is a silent section, no information is sent except for information indicating the segment length. In the case of a consonant part, encoding is performed using the conventional multi-pulse encoding method described in the above-mentioned Documents 1 and 2. In this case, the number of sound source pulses is L ₁ , and the order of the spectrum parameter is M ₁ . Further, when the segmented section is a transient section, encoding is performed using a conventional multi-pulse encoding method. In this case, the order of the spectral parameters is
M ₂ and the number of pulses are L ₂ .

Here, a paper entitled "Effects of vowel stationary and transient parts on clarity in coded speech" by Hanada and Ozawa (Acoustic Society of Japan, 2-1-17, October 1988)
As shown in (Reference 4) and the like, the transient portion is an important section for phoneme perception, and it is necessary to allocate a certain amount of information. Therefore, L ₂ > L ₁ and M ₂ ≧ M ₁ .

The determination of the order of the spectral parameter representing the spectral envelope of the audio signal and the number of multi-pulses representing the audio signal can be performed, for example, as described above.

Also, when the segmented section is a vowel stationary part, the amount of transmitted information may be very small from the viewpoint of phoneme perception. In the vowel stationary part, the speech waveform has a large correlation for each pitch. Therefore, utilizing this correlation, only the segment section length, the pitch period T, and the amplitude correction coefficient (gain) g for each pitch are obtained and transmitted, and the receiving side restores the voice of the vowel stationary part from these information.

The pitch period T and the amplitude correction coefficient g can be specifically determined as follows.

That is, T and g are calculated in accordance with the following equation for each predetermined short section (the number of samples N) shorter than the segment section.

Here, x (n) is the input audio signal of the segment,
(N-T) is a synthesized speech signal encoded and reproduced in an adjacent past segment. w (n) is the impulse response of the perceptual weighting filter.
2 weighting circuit can be referred to. Error power E in short section
In order to obtain T and g that minimize the following equation, the following equation is obtained by partially differentiating equation (1) with respect to g and setting it to 0.

Substituting equation (2) into equation (1) gives Becomes

Here, since the first term of the equation (3) is a constant, the equation (1) is minimized by maximizing the second term. Therefore, a set of g and T that maximizes the second term of Expression (3) may be obtained. This pair of g and T is set to a predetermined short section N
It is calculated and transmitted every time.

Note that this method requires a large amount of computation,
In order to reduce the amount of calculation, the pitch period T may be obtained in advance by the autocorrelation method or the like, and may be substituted into the equation (2) to calculate the value of g. As a method of calculating the pitch period by the autocorrelation method, reference can be made to the pitch extraction circuits of the above-mentioned documents 1 and 2. In order to further reduce the calculation amount, g can be obtained as in the following equation.

Here, P ₁ indicates the power of the synthesized voice signal for one pitch cycle of one past segment, and P ₂ indicates the power of the input voice signal for one pitch cycle of the short section of the corresponding segment. In the method, the characteristics are degraded as the calculation amount is reduced.

Further, as specifically described with reference to FIG. 1, an audio signal is input from an input terminal 100 in the figure, and the input audio signal is stored in a buffer memory 110. The feature extraction circuit 115 obtains feature parameters required for segmentation from the audio signal. As described in the description of the principle, the characteristic parameters are power, pitch, pitch gain, and spectrum change rate. The feature extraction circuit 115 calculates these parameters every 5 msec as described above.

The segmentation circuit 120 inputs a feature parameter, and further uses a temporal change in a 20-30 msec section of the feature parameter in combination with the method described in the principle description,
A segmentation of a silent part, a consonant part, a transient part, and a vowel steady part of the audio signal is performed. Information on the segment section, and when the segment section is a silent section and a vowel stationary section, the segment time length is output from the segmentation circuit 120 to the multiplexer 260.

When the segmented section is other than the silent part and the vowel stationary part, the K parameter calculation circuit 140 performs the well-known LPC analysis on the K parameter as a spectral parameter representing the spectrum characteristic of the audio signal in the segment section, Calculation is performed for a predetermined order. For the specific calculation method, reference can be made to the K-parameter calculation circuits in the above-mentioned documents 1 and 2. The K parameter has the same PARCOR coefficient. Also, the order of the K parameter is, as described above, when the segment section is a consonant part.
M ₁ , and M ₂ (M ₁ ≦ M ₂ ) at the transition part.

K parameter quantization circuit (K parameter encoding circuit)
150 quantizes the K parameter with a predetermined number of quantization bits and outputs the result to the multiplexer 260. Further, this is inversely quantized, converted into a linear prediction coefficient a _i ′, and output.

The impulse response calculation circuit 170 calculates the linear prediction coefficient
Using a _i ′, the impulse response h _w (n) of the synthesis filter subjected to the perceptual weighting is calculated, and this is output to the autocorrelation function calculation circuit 180. The autocorrelation function calculation circuit 180 calculates and outputs an autocorrelation function R _hh (n) of the impulse response up to a predetermined delay time. The operations of the impulse response calculation circuit 170 and the autocorrelation function calculation circuit 180 are described in the aforementioned literature.
1, 2 etc. can be referred to.

The subtracter 190 subtracts the output of the synthesis filter 281 from the audio signal x (n) in the segment, and outputs the subtraction result to the weighting circuit 200. The weighting circuit 200 passes the subtraction result through an auditory weighting filter whose impulse response is represented by _w (n), obtains a weighting signal _xw (n), and outputs it. As described above, the weighting method is described in References 1 and 2.
Etc. can be referred to.

The switch 195 is connected to the terminal 195a (upper part in the figure) when the segment section is a consonant or a transient part, and is connected to the terminal
195b (bottom in the figure).

When the segment section is a consonant or a transient part, the cross-correlation function calculation circuit 210 inputs a weighted signal and a weighted impulse response, calculates a cross-correlation function up to a predetermined delay time, and outputs the result. This calculation method can be referred to the above-mentioned references 1 and 2.

The sound source pulse calculation circuit (sound source signal calculation circuit) 220 obtains and outputs the amplitudes and positions of a predetermined number of multipulses in the segment section. The calculation method of the pulse amplitude and position can be referred to Literatures 1 and 2. The number of pulses is
₁ L when segments consonant as described above, when the transition portion and _two L, and (L ₂ / segment length) ≧ (L ₁ / segment length).

In this way, in the present embodiment, the order obtained by the segmentation is a consonant other than the vowel section, and when the section is a transient part, the order of the K parameter and the number of multi-pulses are determined, and the K parameter and the amplitude of the multi-pulse are determined. Find the position.

The encoding circuit 230 encodes the multi-pulse amplitude g _i and position _mi with a predetermined number of bits and outputs the result. Encoding circuit 23
At 0, these are further decoded and output to the switch 235.

The switch 235 is connected to the terminal 235a (left side in the figure) when the segment section is a consonant or a transient section, and is connected to the terminal 235 when the segment section is a vowel stationary section.
235b (right side in the figure).

The cycle and gain calculation circuit 240 calculates and outputs the cycle T and the gain g for each predetermined short interval according to the equations (1) to (4), as described in the above principle. Here, 5
It shall be calculated every ~ 10meec. Note that an average pitch period can be obtained in a segment section, and T and g can be calculated for each average pitch period.

The encoding circuit 250 quantizes T and g with a predetermined number of quantization bits, and outputs the result to the multiplexer 260. The encoding circuit 250 further decodes and outputs the result to the switch 235.

The synthesis filter 281 receives the output of the switch 235 and, when the segment is a consonant or a transient portion, drives the synthesis filter with a multi-pulse to calculate a synthesized signal of the segment. When the segment is a vowel stationary part, the synthesis filter 281 calculates a short period (5-10
msec), a composite signal is calculated according to the following equation. Note that the receiving side can also synthesize an audio signal in the same manner.

(N) = g · (n−T) (5) Then, the synthesis filter 281 further obtains an influence signal on the next segment caused by the synthesized signal by a predetermined number of samples, and calculates the arithmetic unit 190 Output to In addition, the above-mentioned documents 1 and 2 can be referred to for the calculation method of the influence signal.

As described above, the audio signal is classified into several segments based on the phonetic features, and the amount of information necessary for human phonological perception is assigned and transmitted according to the segment. Since the amount of transmitted information is greatly reduced and speech is restored from past segments using waveform correlation, the bit rate can be reduced to 4.8 kb / s, which is important for the perception of phoneme and naturalness Also, a synthesized voice with good sound quality can be obtained in a portion where the characteristics of a proper voice are changed (voiced transition portion or a portion between vowels).

Note that the above-described embodiment is merely one configuration of the present invention, and various modifications thereof are also conceivable.

That is, in the embodiment, the audio signal of the segment is classified into a vowel steady part, a consonant part, a transient part, and a silent part, but the number of classifications may be changed. Further, the consonant part may be further classified into a friction part and a rupture part.

Further, in the embodiment, the K parameter is encoded as a spectrum parameter, and LPC analysis is used as an analysis method, but other known parameters such as LSP, LPC cepstrum, cepstrum, and the like are used as the spectrum parameter.
Improved cepstrum, generalized cepstrum, mel cepstrum and the like can also be used. In addition, an analysis method suitable for each parameter can be used.

When the segment is a vowel stationary part, the pitch period T and the gain g are calculated based on the past segment. However, the pitch period T and the gain g may be calculated from the past and future segments sandwiching the segment. it can. This method is described in, for example, a paper by Ikeda and Itakura entitled "Encoding of Speech Waveform Using Block Decimation and Interpolation" (Acoustic Society of Japan Electroacoustics Research Society, EA88-15, 1988) (Reference 5). it can. However, in this method, the amount of transmission information increases as compared with the method of the embodiment.

Further, the pitch period T and the gain g are set to the first order, but may be set to a higher order. For example, assuming a third order, equation (5) can be written as follows.

(N) = g ₁ · (n−T−1) + g ₂ · (n−T) + g ₃ · (n−T−1) (6) By setting higher order, the sound quality of the vowel stationary part But the amount of transmitted information increases.

Also, in connection with all the methods described above, instead of sending the pitch cycle as it is for each short section, an average pitch cycle is obtained and sent in the segment section, and the average pitch cycle and the pitch cycle obtained for each short section are sent. The difference d from T may be transmitted. This can reduce the amount of transmission information required for transmission of the pitch period T.

Also, in the segment of the vowel stationary part, the voice signal of one pitch section is expressed by a multi-pulse and a spectrum parameter for each predetermined section and transmitted, and the above-described pitch period and gain are transmitted in other sections. You may.
By doing so, the sound quality of the vowel stationary part is further improved, but the amount of transmitted information increases.

As is well known in the field of digital signal processing, since the autocorrelation function corresponds to the power spectrum on the frequency axis and the cross-correlation function corresponds to the cross power spectrum, it can be calculated from these. For a discussion of these calculations, see “Digital Sig
nal Processing "(Prentice-Hall, 1975).

〔The invention's effect〕

As described above, according to the present invention, it is possible to classify an audio signal into several segments based on feature parameters, allocate a necessary amount of information according to those segments, and transmit the information. It is possible to easily reduce the bit rate and reduce the bit rate, and a large effect that it is possible to obtain a synthesized voice with good sound quality even in the part where the characteristics of the voice that is important for phonological perception and perception of naturalness changes. is there.

[Brief description of the drawings]

FIG. 1 is a block diagram showing one embodiment of a speech encoding method according to the present invention. 100 input terminal 110 buffer memory 115 feature extraction circuit 120 segmentation circuit 140 K parameter calculation circuit 150 K parameter coding (K parameter quantization circuit) 170 impulse response calculation circuit 180 …… Autocorrelation function calculation circuit 195,235… Switch 195a, 195b, 235a, 235b… Terminal 200… Weighting circuit 210… Cross correlation function calculation circuit 220… Sound source pulse calculation (sound source signal calculation) circuit 230,250… Encoding circuit 240: Period and gain calculation circuit 260: Multiplexer 281: Synthesis filter 300: Output terminal

Continuation of the front page (58) Field surveyed (Int. Cl. ⁶ , DB name) G10L 3/00-9/18 H03M 7/30 H04B 14/04 JICST (JOIS)

Claims (1)

(57) [Claims] 1. A feature parameter representing a feature of a voice signal is obtained from an input discrete voice signal, the voice signal is segmented in accordance with the feature parameter, and a section obtained by the segmentation is a section other than a vowel section. When the order of the spectrum parameter representing the spectrum envelope of the audio signal of the section and the number of multi-pulses representing the sound source signal of the audio signal are determined, the spectrum parameter and the amplitude and position of the multi-pulse are obtained, and obtained by the segmentation. When the obtained section is a vowel section, a coefficient for predicting a signal of the vowel section is obtained from a signal of at least one section adjacent to the vowel section.