JP3055901B2 - Audio signal encoding / decoding method and audio signal encoding device - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an audio signal encoding method and an audio signal encoding device for efficiently encoding an audio signal at a low bit rate. The present invention relates to a speech signal encoding method capable of non-uniformly dividing speech based on a parameter and obtaining and encoding a parameter representing a feature of the speech signal in the divided section, and an apparatus used therefor.

[Conventional technology]

As a method of transmitting an audio signal at a low transmission bit rate, for example, 8 kb / s or less, a pitch prediction multi-pulse encoding method at about 8 kb / s and a pitch interpolation multi-pulse encoding method at about 4.8 kb / s are known. ing. In each of these, the sound source signal is a combination of multiple pulses (multi-pulse)
, The characteristics of the vocal cords are represented by a digital filter, and the information of the sound source pulse and the coefficient of the filter are obtained and transmitted for each fixed time section (frame). For details of this method, the former is described, for example, by “High Qualit” by Ozawa, Araseki.
y Multi-pulse Speech Coder with Pitch Prediction "
(Proc.ICASSP, Lecture number 33.3 1986) (Reference 1)
For the latter, for example, “Low” by Ozawa and Araseki
Bit Rate Multi-pulse Speech Coder with Natural S
peech Quality "(Proc. ICASSP, Lecture number 9.7,198
6) (Reference 2). With these methods,
In order to reduce the amount of transmission information, the source pulse information to be transmitted is reduced by predicting the pitch of the source pulse signal or obtaining a pulse train only for one pitch section in the frame.

[Problems to be solved by the invention]

However, in this conventional method, the section length for obtaining the sound source pulse and the filter coefficient is fixed (20 ms in References 1 and 2).
And had Therefore, in the vowel section, although a substantially periodic waveform continues and the characteristics of the voice have not changed much,
Since information is transmitted every short time of 20 ms, the efficiency is extremely low. On the other hand, there is a problem in that consonant sections cannot follow changes in characteristics of fast speech and sound quality deteriorates. In particular, this problem was remarkable when the bit rate was considerably lower than 8 kb / s.

To explain the above problem more specifically, first, as is well known, a vowel section has a long time length, generally 100 to 300 ms, depending on the generation speed. It can be regarded as a stationary section where the feature hardly changes.
Furthermore, it is known that, in the vowel stationary part, even if the signal is suppressed to zero and no information is transmitted, the syllable clarity hardly deteriorates. However, naturalness deteriorates. Therefore, as in the conventional method, if the information is transmitted by analyzing it for each short frame section of about 20 ms, the efficiency is very low. On the other hand, in the consonant section, the characteristics of the voice change rapidly, so
Frame is too long, it is difficult to perform accurate analysis corresponding to a change in the voice, and the sound quality of the reproduced voice has deteriorated.

Therefore, in order to improve these problems, for example, Ma
"Linear Prediction of Speech" by rkel, Gray No. 1
As described in Chapter 0 (Springer-Verlag, 1976) (Reference 3), the frame length is set to an integral multiple of the fixed interval based on the change in the difference between the frames of the spectrum obtained with the fixed-length frame of about 10 ms. However, there is a problem with this improvement. That is, in such a method, the frame length is changed using the characteristic parameter that is not well associated with the auditory sense in the above manner, and the method of changing the frame length is based on the fixed section length. Because of the lack of flexibility, if the bit rate is reduced by increasing the frame length increasing section, there is a problem that the sound quality is greatly deteriorated.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a speech signal coding method capable of greatly reducing the amount of information required for transmitting a sound source signal, and capable of greatly reducing auditory deterioration of synthesized speech even if the bit rate is significantly reduced. It is an object of the present invention to provide a method and a speech signal encoding device.

The speech signal encoding / decoding method of the present invention includes the steps of: inputting a discrete speech signal; obtaining a cepstrum parameter for each predetermined time interval; calculating a scale indicating a time change of the cepstrum parameter; The audio signal is divided into non-uniform time sections by searching for a time position near the maximum time position and at which the feature amount falls below a predetermined threshold and determining the boundary as a segment boundary, The sound source signal in all or a part of the divided section is represented by a combination of a plurality of pulse trains and transmitted, and the sound source signal in the section is restored using the transmitted pulse train to represent the audio signal. It is characterized by outputting an audio signal.

Further, the speech signal encoding apparatus of the present invention obtains a cepstrum parameter for each predetermined time interval from the input discrete speech signal sequence, and calculates a segmentation scale calculation for calculating a scale indicating a time change of the cepstrum parameter. A circuit for searching for a time position near the time position at which the scale is maximal and at which the scale is smaller than a predetermined threshold value and determining the segment position as a segment boundary, thereby making the voice signal a non-uniform time interval. A segmentation circuit that performs segmentation on all of the divided sections; a spectrum parameter calculation circuit that calculates a spectrum parameter and a pitch parameter representing a short-time spectrum characteristic from all or some of the divided sections; and all or one of the divided sections. Of the sound source signal in the section A sound source pulse calculating circuit for calculating a combination of the scan row, characterized by comprising a multiplexer circuit for outputting a combination of the spectral parameters and pitch parameters and excitation pulse train.

[Operation] Since the audio signal is encoded as described above, the amount of information necessary for transmission of the sound source signal can be significantly reduced, and even if the bit rate is significantly reduced, the visual deterioration of the synthesized voice is extremely small. An encoding process can be performed.

The audio signal encoding device includes the segmentation circuit, the spectrum parameter calculation circuit, the excitation pulse calculation circuit, and the multiplexer circuit configured as described above, so that the above-described encoding process can be performed.

〔Example〕

Next, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing a configuration of an embodiment of an audio signal encoding method and an audio signal encoding device according to the present invention. FIG. 1 also shows an audio signal decoding device to facilitate understanding of the invention. FIG. 2 is a block diagram for explaining the principle of audio signal encoding.

In the encoding / decoding method of the audio signal according to the embodiment of FIG. 1, the transmitting side inputs a discrete audio signal, and divides the audio signal into non-uniform sections by a method that is compatible with the auditory characteristics, The sound source signal in all or part of the divided sections is represented by a combination of a plurality of pulse trains and transmitted.
On the receiving side, the sound source signal in the section is restored by using the pulse train, and a synthesized voice signal that well represents the voice signal is output.

Hereinafter, first, the principle of the encoding process according to the present invention will be described with reference to FIG. In the figure, a segmentation scale calculation unit 400 inputs an audio signal,
For each short time period (for example, 5 ms) in which a consonant part having a rapid change in speech characteristics can be accurately analyzed, a scale that has a good correspondence with the auditory characteristics and well represents the time change of the characteristics of the speech signal is used as a segmentation measure. Here, the following dynamic scale D (t) is used as this scale. First, an LPC cepstrum Ci (1) is used as a parameter that expresses a good spectrum envelope from a speech signal every short time (for example, 5 ms).
.Ltoreq.i.ltoreq.p), which is converted into a dynamic measure based on equation (1). Here, the dynamic scale is a parameter known in the field of speech recognition, and satisfactorily represents a temporal change in a phoneme feature amount of a speech signal when captured by hearing. Generally, a portion having a large temporal change in a phoneme of a voice signal (a consonant or a transition between a consonant and a vowel) has a large value, and a stationary portion (a vowel stationary portion or the like) has a small value.

Where a _i is It is.

For a detailed explanation of this calculation method, see "On the Role of Spectral Transition for Speech" by Furui.
Peception "(J. Acoustical Society of Am
erica, vol. 80, pp. 1016-1025, 1986) (Reference 4), and the details are omitted here. Also,
(1) Power section a ₀ were included (3) instead of expression can also be used. By including the power term by equation (3),
It is possible to consider not only the spectral feature but also the time change of the power feature, and it is possible to improve the accuracy of the scale.

The segmentation unit 410 calculates the dynamic scale D (t)
To divide the audio signal non-uniformly (segmentation). This is performed using the scale of the above equation (1) or (3). First, an audio signal is divided in advance around each local maximum value of the scale. Here, as described in the division 4, the part of several tens ms before and after the maximum value of the scale substantially corresponds to the articulatory connection part from a consonant to a vowel and from a vowel to a consonant. Has been reported to be a very important part of the hearing. Therefore, the audio signal is segmented at a place where the scale becomes smaller to some extent continuously except for such an auditory important part. FIG. 2B shows the state of the segmentation. Here, the upper part of FIG. 2B shows an audio waveform, and the lower part shows an example of a dynamic scale and a segmentation.

Next, the LPC / pitch analysis unit 430 analyzes the audio signal of the entire segmented section or a part of the segmented section to obtain an LPC coefficient. In the case of obtaining from a part of the audio signal, the cepstrum obtained by the segmentation unit 410 can be converted into LPC coefficients by a surrounding method. Then, the pitch period is calculated by a known method, and it is determined whether or not the segmented section is a vowel stationary part. Here, for this determination, a method of determining whether the value of the autocorrelation function (pitch gain) separated from the power in the segmentation section by the pitch period is larger than a predetermined threshold value can be used. .

When the segmented section is a vowel stationary section, the sound source calculation section 420 divides the segmentation section into subframes for each period of the pitch section, and calculates a sound source pulse train for one of the pitch sections.
The calculation of the sound source pulse train can be referred to Japanese Patent Application No. 59-272435 (Reference 5).

Further, for other pitch sections, an amplitude correction coefficient is obtained for each pitch section so that the waveform of the pitch section is well represented.

Therefore, according to the present invention, it is possible to greatly increase the number of excitation pulses in one pitch section even if the bit rate is significantly reduced as compared with the conventional method. Alternatively, even if the reconstruction is performed using interpolation processing, the sound source signal of the entire section can be satisfactorily represented.

On the other hand, if the segmentation section is not a vowel stationary section, a sound source pulse train is obtained for the entire section.

The transmission information on the transmitting side is the amplitude and position of the sound source pulse train, segmentation information indicating the length of the segmented section, pitch period, discrimination information, and amplitude correction coefficient. On the receiving side, at the time of the vowel stationary part, the amplitude and position of the transmitted sound source pulse train are smoothly changed for each pitch cycle, or the sound source signal between the segmented sections is subjected to interpolation processing and transmitted. The pulse train in the pitch section other than the pitch section is restored, and the sound source signal in the segmented section is restored.

 Next, a description will be given with reference to FIG.

In FIG. 1, the transmitting side includes an audio signal encoding device, and the receiving side includes an audio signal decoding device, and an appropriate transmission path is provided between the two.

An audio signal encoding device, a segmentation circuit that extracts a characteristic parameter corresponding to the identification of the auditory sense from the input discrete audio signal sequence and segments the audio signal into non-uniform time intervals using the parameter, A spectrum parameter calculation circuit for calculating a spectrum parameter and a pitch parameter representing a short-time spectrum characteristic from the divided voice signal, and a plurality of pulse trains representing a sound source signal in all or some of the divided sections. A sound source pulse calculation circuit for calculating the combination; and a multiplexer circuit for combining and outputting the spectrum parameter, the pitch parameter, and the sound source pulse train.

An audio signal decoding apparatus receives a spectrum parameter, a pitch parameter, and an excitation pulse train representing an excitation signal, representing a short-time spectrum characteristic of an audio signal, and separates the spectrum parameter, the pitch parameter, and the excitation pulse train from each other. A sound source restoring circuit for restoring a sound source signal of an entire section which is non-uniformly divided using the pitch parameter and the sound source pulse train; and a synthesis section for synthesizing a sound signal of the section using the restored sound source signal. And a filter.

The audio signal encoding and decoding processes are performed as follows.

FIG. 1 shows an embodiment of the present invention.
Input a discrete audio signal from 0. The segmentation scale calculation circuit 505 performs the same calculation as the segmentation scale calculation section 400 in FIG. 2A, and outputs a segmentation scale. The segmentation circuit 510 performs the same processing as the segmentation unit 410 in FIG. 2 (a), segments the audio signal into non-uniform sections, and outputs segmentation information indicating the length of the segmentation section and the segmented audio signal. I do. The LPC / pitch calculation circuit 520 performs the same processing as the LPC / pitch analysis unit 430 in FIG. 2 (a), and performs LPC analysis, pitch period calculation and segmentation of the vowel stationary section on the segmented voice signal. It determines whether or not it is a part, and outputs the LPC coefficient, the pitch period, and the determination information to the quantizer 530. The quantizer 530 quantizes the information with a predetermined number of bits, outputs the quantized information to the multiplexer 600, and dequantizes the information.

The weighting circuit 540 weights the segmented audio signal using the dequantized LPC coefficient and the signal. The weighting method can be referred to the weighting circuit (200) of the above-mentioned reference 5. The impulse response calculation circuit 560 calculates an impulse response using the inversely quantized LPC coefficients. For the method of calculating the impulse response, reference can be made to the impulse response calculation circuit (170) of Reference 5. The autocorrelation function calculation circuit 570 calculates the autocorrelation function of the impulse response and outputs the calculated autocorrelation function to the sound source pulse calculation circuit 580. The calculation method of the autocorrelation function can be referred to the autocorrelation function calculation circuit (180) of Reference 5 described above. The cross-correlation function calculation circuit 550 calculates a cross-correlation function between the weighted signal and the impulse response, and outputs the result to the sound source pulse calculation circuit 580. For this calculation method, reference can be made to the cross-correlation function calculation circuit (210) of the above-mentioned Document 5.

When the segmented section is a vowel stationary section, the sound source pulse calculation circuit 580 divides the section into subframes for each pitch period and converts the section into a vicinity of the center, as described in the description of FIG. The excitation pulse train is calculated for the subframe section of. In other sub-frame sections, the amplitude correction coefficient of the pulse train is obtained one by one in each section as described in the description of FIG. 2A. On the other hand, when it is not a vowel stationary part, a sound source pulse train is calculated for the entire section. For the method of calculating the sound source pulse train, the drive signal calculation circuit (220) of Reference 5 can be referred to. The quantizer 590 quantizes the amplitude and position of the sound source pulse train with a predetermined number of bits and outputs the result to the multiplexer 600. The operation of the quantizer 590 can be referred to the encoding circuit (230) of the aforementioned document 5. The multiplexer 600 combines and outputs the quantized sound source pulse train, LPC coefficient, pitch period, segmentation information, discrimination information, and amplitude correction coefficient.

On the receiving side, on the other hand, the demultiplexer 610 separates and outputs the excitation pulse information, the LPC coefficient, the pitch period, the segmentation information, the discrimination information, and the amplitude correction coefficient. The excitation pulse decoder 620 decodes the amplitude and position of the excitation pulse train. The LPC / pitch decoder 650 decodes the LPC coefficient and the pitch period. The sound source restorer 630 receives the discrimination information and the segmentation information, and when the section is a vowel stationary part, restores and outputs the sound source signal of the entire segmentation section using the decoded one-pitch section sound source pulse train. Here, as a method of restoring the sound source pulse train in the pitch section that is not transmitted, the position is restored by shifting the entire pulse in the pitch section by the pitch period, and the amplitude is restored by multiplying by the amplitude correction coefficient. In addition to this method, a method of restoring by an interpolation process using a sound source pulse train in an adjacent segmentation section is known, and the details can be referred to the above-mentioned reference 5. In addition, other well-known methods can be used. On the other hand, when the section is not a vowel stationary section, a sound source signal for the entire section is generated and output using the received sound source pulse train. Interpolator 640
Using the decoded LPC coefficient, discrimination information, and pitch period,
When the segmentation section is a vowel stationary part, the LPC coefficient is interpolated on the PARCOR coefficient for each pitch cycle in order to smooth the spectrum change. On the other hand, when the section is not the vowel stationary part, the coefficient is output to the synthesis filter 660 without interpolation. This is to prevent a large distortion from being generated by interpolation because the spectral characteristics of the audio signal change rapidly in portions other than the vowel stationary part. The synthesis filter 660 synthesizes the audio signal in the entire segmentation section using the LPC coefficient, the restored sound source signal, and the segmentation information, and outputs the synthesized signal through the terminal 670.

As described above, according to the above configuration, the audio signal is non-uniformly segmented using the characteristic parameters that are well associated with the characteristics of the auditory sense, and a plurality of types of vector quantum When the section is a vowel stationary part that is long in terms of time with little change in speech characteristics, the sound source pulse train is generated for one pitch section of the section. Since the sound source pulse train is obtained for the entire section except for the vowel stationary part, the amount of information necessary for sound source signal transmission can be greatly reduced. Therefore, even if the bit rate is greatly reduced, the perceived deterioration of the synthesized speech is very small and high naturalness can be obtained.

The above-described embodiment is merely an embodiment of the present invention, and various modifications thereof can be considered.

For example, when the segmented section is a vowel stationary section, the cross-correlation function calculation circuit 550 calculates the cross-correlation function only for one pitch section near the center of the section, not for the entire section. Is also good. This is because the sound source pulse train is actually obtained in one pitch section. In this method, the characteristics are slightly deteriorated, but the amount of calculation can be reduced to almost P / N (where P is the pitch period, and N is the length of the segmentation section of the vowel stationary part).

As a method of calculating the sound source pulse, a known good method can be used in addition to the above-described embodiment. See K. Ozawa, “A Study of Pulse Search Algorithms.
for Multi-pulse Speech Codec Realization "(J. Sele
cted Area of Communications, pp., 1987) (Reference 6).

When the segmented section is a vowel stationary section, the position of one pitch section for obtaining a sound source pulse train is not fixed, but is searched for and determined so as to obtain the best synthesized speech. You can also. This process further improves the sound quality but slightly increases the amount of calculation. Reference 5 can be referred to for a specific method.

As a method of interpolating the coefficients of the synthesis filter 660, interpolation can be performed on the logarithmic cross-sectional area ratio or other parameters. Further, as an interpolation method, logarithmic interpolation or the like can be used in addition to linear interpolation. See B. for details on these methods.
“Speech Analysis and Synthesis by L by S. Atal
inear Prediction of the Speech Wave "(J.Acoust.So
c. America, pp. 637-655, 1971) (Reference 7).

Also, the synthesized sound quality is further improved by smoothly changing the pitch period on the receiving side by interpolation.

[Effects of the Invention] As described above, according to the audio encoding / decoding method of the present invention, when an audio signal is encoded and decoded by decoding, a synthesized or audio signal that satisfactorily represents the audio signal is generated. Obtainable. For this reason, deterioration in sound quality can be reduced as compared with the conventional fixed-length frame or the one in which the frame length is changed based on the change in the difference between the frames of the spectrum obtained in the fixed-length frame. .

Further, according to the present invention, a sound quality signal can be divided into non-uniform sections using feature parameters that are well associated with auditory characteristics. As a result, the divided sections have a long vowel with little change in speech characteristics. In the case of a stationary part, one of the sections
A sound source pulse train is obtained for one pitch section, and a sound source pulse train can be obtained for the entire section other than the vowel stationary section, so that the amount of information required for transmission of a sound source signal can be greatly reduced.

Therefore, according to the present invention, even if the bit rate is significantly reduced, it is possible to perform the encoding / decoding processing which can obtain a very natural sound with very little auditory deterioration of the synthesized speech.
This method is suitable for efficiently encoding and decoding an audio signal at a low bit rate.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an embodiment of an audio signal encoding method and an audio signal encoding device according to the present invention, and FIG. 2 is a principle block diagram and a waveform diagram for explaining the present invention. 400 segmentation scale calculation unit 410 segmentation unit 420 sound source calculation unit 430 LPC, pitch analysis unit 505 segmentation scale calculation circuit 510 segmentation circuit 520 LPC pitch calculation circuit 530,590 quantum Transformer 540 Weighting circuit 550 Cross-correlation function calculation circuit 560 Impulse response calculation circuit 570 Autocorrelation function calculation circuit 600 Multiplexer 610 Demultiplexer 620 Decoder 630 Sound source restoration 640 …… Interpolator 650 …… LPC, pitch decoder 660 …… Synthesis filter

Claims (2)

(57) [Claims] 1. A discrete audio signal is input, and cepstrum parameters are determined for each predetermined time interval.
Calculating a scale indicating the temporal change of the cepstrum parameter, searching for a time position near the time position at which the scale is maximal and at which the scale falls below a predetermined threshold value, and determining the boundary as a segment boundary. Divides the audio signal into non-uniform time sections, and transmits the sound source signal in all or some of the divided sections by expressing a combination of a plurality of pulse trains, and using the transmitted pulse trains. An audio signal encoding / decoding method for restoring a sound source signal in the section and outputting a synthesized audio signal representing the audio signal. 2. A segmentation scale calculation circuit for obtaining a cepstrum parameter at predetermined time intervals from an input discrete audio signal sequence, and calculating a scale indicating a time change of the cepstrum parameter; A segmentation circuit for segmenting the audio signal into non-uniform time intervals by searching for a time position near the time position to be and where the scale is less than a predetermined threshold and determining it as a segment boundary, A spectrum parameter calculation circuit that calculates a spectrum parameter and a pitch parameter representing a short-time spectrum characteristic from all or a part of the divided sections; and a sound source signal in all or part of the divided sections. Calculate multiple pulse train combinations to represent That the sound source pulse and calculating circuit, an audio signal encoding apparatus and a multiplexer circuit for outputting a combination of the spectral parameters and pitch parameters and excitation pulse train.