JP3490324B2 - Acoustic signal encoding device, decoding device, these methods, and program recording medium - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a high-efficiency acoustic signal encoding method for digitally encoding a signal sequence of an acoustic signal such as voice with a small amount of information, a decoding method therefor, and these devices.

[0002]

2. Description of the Related Art In digital mobile communication, a high-efficiency voice coding method is used in order to efficiently use radio waves and to efficiently use communication lines and storage media for voice or music storage services. . At present, as a method for efficiently encoding speech, original speech is called a frame or subframe (hereinafter collectively referred to as frame) for 5 to 50 m.
It is divided into sections at a constant interval of about s, and the sound of one frame is separated into two pieces of information, a characteristic of a linear filter that represents the envelope characteristic of the frequency spectrum and a driving sound source signal for driving the filter, A method of encoding each has been proposed. In this method, the pitch period (fundamental frequency) of the voice is used as a method of encoding the driving excitation signal.
A method is known in which a periodic component that is considered to correspond to the above and a component other than that are encoded separately. Code-Excited Linear Prediction (CELP) is an example of a method of encoding the drive excitation information. For more information on this technique, see the document MR Schroeder a
nd BS Atal, “Code-Excited Linear Prediction (C
ELP): High Quality Speech at Very Low Bit Rate
s ", IEEE Proc. ICASSP-85, pp.937-940, 1985.

FIG. 3 shows an example of the configuration of the above encoding method. The speech x input to the input terminal 11 is the linear prediction analysis unit 1
In 2, the linear prediction parameter a representing the frequency spectrum envelope characteristic of the input speech is calculated. The obtained linear prediction parameter a is the linear prediction parameter encoding unit 13
In, the quantization and coding are performed, and the quantized value is converted into the synthesis filter coefficient aq and sent to the synthesis filter 14, and the code ba is sent to the code sending unit 15. When the distortion calculation is performed by using the spectral information of the input voice such as considering the auditory characteristics in the distortion calculation, the linear prediction parameter a or the quantized linear prediction parameter aq is also input to the waveform distortion calculation unit 16. Sent. Details of the linear prediction analysis and an example of coding the linear prediction parameters are described in, for example, “Digital Speech Processing” by Sadahiro Furui (Tokai University Press). Here, the linear prediction analysis unit 12,
Linear prediction parameter coding unit 13 and synthesis filter 1
4 may be replaced with a non-linear one.

The driving sound source vector generation unit 17 generates driving sound source vector candidates having a length of one frame and sends them to the synthesis filter 14. The drive excitation vector generation unit 17 is generally composed of an adaptive codebook 18 and a fixed codebook 19. From the adaptive codebook 18, the immediately preceding past drive excitation vector (the already quantized drive excitation vector for one to several frames immediately before) stored in the buffer is cut out at a length corresponding to a certain cycle, and the cutout is performed. By repeating the above vector until the length of the frame is reached, the candidates of the time series vector corresponding to the periodic component of the voice are output. The "certain cycle" is selected as a cycle in which the distortion in the waveform distortion calculation unit 16 is small, but the selected cycle is generally often equivalent to the pitch cycle of voice. The fixed codebook 19 outputs candidates for a time-series code vector having a length of one frame, which corresponds to the aperiodic component of speech. Depending on the number of bits for encoding, these candidates are independent of the input speech and are stored in a predetermined number of candidate vectors, and are one of them.
It may be one of the vectors generated by arranging the pulses according to a predetermined generation rule. The fixed codebook 19 originally corresponds to a non-periodic component of speech, but particularly in a speech section having a strong pitch periodicity such as a vowel section, the previously prepared candidate vector is
A fixed code vector may be obtained by applying a comb filter having a pitch period or a period corresponding to the pitch used in the adaptive codebook, or by cutting out and repeating the vector as in the process in the adaptive codebook. The time-series vector candidates ca and cr output from the adaptive codebook 18 and the fixed codebook 19 are multiplied by gain candidates ga and gr output from the gain codebook 23, respectively, in the multiplication units 21 and 22, and added by the addition unit. At 24, they are added and become the drive sound source vector candidate c. In the configuration example of FIG. 3, only the adaptive codebook 18 or the fixed codebook 19 may be used during the actual operation.

The synthesis filter 14 is a linear filter having a synthesis filter coefficient obtained from the linear prediction parameter aq quantized in the linear prediction parameter coding unit 13 as a filter coefficient, and receives the driving sound source vector candidate c as an input. The candidate y of the reproduced voice is output. As the order of the synthesizing filter 14, that is, the order of the linear prediction analysis, the order of 10 to 16 is generally used. In addition,
As already mentioned, the synthesis filter 14 may be a non-linear filter.

In the waveform distortion calculator 16, the synthesis filter 1
The distortion d between the reproduced voice candidate y which is the output of No. 4 and the input voice x is calculated. The calculation of this distortion is often performed in consideration of the coefficient aq of the synthesizing filter 14 or the non-quantized linear prediction coefficient a, as represented by auditory weighting. The codebook search control unit 25 selects the drive excitation code bc, that is, the periodic code, the fixed (noise) code, and the gain code such that the distortion between each reproduced speech candidate y and the input speech x is the minimum or equivalent to the minimum. Determine the driving source vector in the frame.

The driving excitation code bc (periodic code, fixed code, gain code) determined by the codebook search control unit 25 and the linear prediction parameter code ba output from the linear prediction parameter coding unit 13 are the code sending unit. 15 and is stored in the storage device or sent to the receiving side via the communication path depending on the form of use. FIG. 4 shows a configuration example of a decoding method corresponding to the above encoding method. Among the codes received by the code receiving unit 31 from the transmission path or the storage medium, the linear prediction parameter code ba is the synthesis filter coefficient aq in the linear prediction parameter decoding unit 32.
And is sent to the post-processing section (also called a post filter) 34 if necessary. Of the received codes, the drive excitation code bc is sent to the drive excitation vector generation unit 35, and the excitation vector c corresponding to the code is generated. The synthesis filter 33 receives the driving sound source vector c as an output and outputs a synthesized speech y. The post-processing unit 34 subjects the synthesized speech y to spectrum enhancement and pitch enhancement processing to aurally reduce quantization noise. Since the post-processing unit 34 is a kind of voice enhancement process, it may not be used depending on the relation of the processing amount and the characteristics of the input signal. The driving excitation vector generation unit 35 selects the time series vector ca from the adaptive codebook 36 according to the periodic code in the driving excitation code bc, and also selects the time series vector cr from the fixed codebook 37 according to the fixed code. ca and cr are multiplication units 38 and 39
Then, the gains g a and g r extracted from the gain code book 41 are multiplied by the gain code, added together in the adder 42, and input to the synthesis filter 33 as the driving sound source vector c. As described above, when only the adaptive codebook 18 or only the fixed codebook 19 is used for encoding during the actual operation, correspondingly, only the adaptive codebook 36 or the fixed codebook 37 in FIG. Is used.

[0008]

A problem with a coding method based on a speech generation model such as such a CELP-based coding method is that a speech signal recorded in a quiet environment and having no background noise is a problem. When is input, high-quality encoding can be achieved with a small amount of information, but when a voice recorded in an environment with background noise, such as an office or the street, is input, such as curcules and crackles. That is, a very unpleasant sound is reproduced.
The problem when these background noises are input is that the CELP-based speech coding model using pitch periodicity is based on a speech generation model, whereas background noise exhibits different properties from speech. Is. Specifically, while the adaptive codebook outputs a signal component corresponding to the pitch period of speech, background noise generally has no pitch periodicity, so an unnatural periodic sound is generated in the background noise section. To do. Also, in the voice section where background noise is superimposed,
Originally, although it is a signal of the nature that a voice signal with pitch periodicity and a noise signal without pitch periodicity are added,
In order to apply the coding model that emphasizes the pitch periodicity of speech, the background noise component is also superposed as an unnatural periodic sound. When the fixed codebook is used with a pitch period, the pitch period of the fixed codebook also causes an unnatural periodic sound. As described above, when the configuration of the adaptive codebook or fixed codebook does not match the characteristics of the signal,
There is also a problem in the method of determining the gain. That is, the conventional gain determination method is based on the premise that the characteristics of the driving excitation vector output from the adaptive codebook or the fixed codebook match the characteristics of the input signal well. When the property does not match the property of the input signal, the conventional method results in an unnaturally varying signal.

As a typical solution to this problem,
There are a method using noise reduction and a method called comfort noise generator. In the former method, noise reduction processing is performed before the input signal to relatively reduce the background noise component, and the unpleasant sound in the reproduced sound is reduced by the amount of the noise component reduced. However, the noise reduction process does not completely eliminate the noise, and the unpleasant noise cannot be eliminated. Further, when the background noise is a non-stationary sound, it is difficult to obtain a sufficient noise reduction effect itself. On the other hand, the latter comfort noise generator directly encodes the voice section using the CELP system encoding method, and generates and replaces appropriate "comfortable" noise, such as white noise, for the noise section. If you use the comfort noise generator method, unpleasant sounds such as curcules will not be played, but the noise interval of the sound that is played will always be the same as the nature of background noise such as offices and crowds. As a result, the problem that the background sound information is not transmitted to the receiving side occurs. Also, when the level of background noise is high, it is difficult to switch between the voice section and the noise section without error, and the reproduced sound is deteriorated due to the section detection error, or even if there is no section detection error. There were too many differences in the nature of the noise section, and it often sounded discontinuous.

According to the present invention, in a voice encoding system based on a voice generation model, such as a CELP system, an unpleasant sound is not reproduced and the nature of the background sound is transmitted to the receiving side so that it is more natural. An object of the present invention is to provide a method and apparatus for encoding and decoding that realizes various reproduced sounds.

[0011]

According to the present invention, the characteristics of background noise are defined as short-term characteristics (short-term fluctuation component) such as rattling noise, vehicle passing sound, footsteps, and human voice in the distance. For example, the two characteristics of an average long-term characteristic (long-term fluctuation component) such as a steady noise and a rotating sound of a motor are regarded as the transmission parameters in the framework of the CELP system coding model. Information is carried and sent to the receiving side, a signal that reproduces both characteristics is generated on the reproducing side, and by mixing them, a natural reproduced sound is output even in the case of voice input with a background sound. The point of the invention is that the noise characteristics are well transmitted without significantly switching the framework of the coding model specialized for speech and in the limited amount of information (bit rate). The detection of the section does not need to be so strict.

According to the decoding method of the present invention, the adaptive codebook and / or the fixed codebook or both of them are used as a frame unit or a subframe unit (hereinafter collectively referred to as a frame unit).
The code vector extracted in step 1 is multiplied by the gain extracted from the gain codebook to generate a driving sound source vector, and a driving filter is driven by the driving sound source vector to generate a voice signal or an acoustic signal (hereinafter collectively referred to as a voice signal). In the method of decoding the generated voice, receiving information on whether or not the corresponding frame is in the background noise section, measuring the power of the output signal of the synthesis filter in the frame in the background noise section,
The power level representing the long-term average value is calculated, and in the frame in the background noise interval, the average spectrum representing the long-term average of the spectral parameters representing the filter coefficient of the synthesis filter is calculated, and the characteristics of the average spectrum are represented. The signal generated by driving the filter with white noise is
Amplitude adjustment based on the measured power level, to generate a background noise stationary component signal, the generated background noise stationary component signal, regardless of whether the frame is a background noise section, It is added to the output signal of the synthesis filter to generate a reproduced voice.

Further, a stationary sound having a constant spectral characteristic and a constant amplitude independent of the characteristic of the input signal is generated together with a signal generated by driving a filter showing the characteristic of the average spectrum with white noise, and the background noise. Is added to the output signal of the synthesizing filter as a stationary component signal of, and reproduced sound is generated. It also receives information on whether the frame is in the background noise section or consonant section, measures the power of the driving sound source vector for each frame, and uses the amplitude determined based on the power of the driving sound source vector as the white noise. The background noise fluctuation component signal generated by multiplication is added to the driving sound source vector when the frame is in the background noise section or the consonant section.

According to the encoding method of the present invention, a code vector extracted in a frame unit or a subframe unit (hereinafter collectively referred to as a frame unit) from both or one of the adaptive codebook and the fixed codebook is used by the gain codebook. Multiply the extracted gain to generate a driving sound source vector,
In an encoding method for selecting an adaptive code, a fixed code, or a gain code by comparing an audio signal or an audio signal generated by driving a synthesis filter (hereinafter collectively referred to as an audio signal) with an input audio signal, Analysis, the frame is
It is determined whether or not it corresponds to the background noise section or the consonant section, and when the frame corresponds to the background noise section or the consonant section, the distance based on the waveform distortion minimization of the output signal of the synthesis filter and the input signal. By using a weighted sum of the scale and the distance measure based on the power level difference minimization of the output signal and the input signal of the synthesis filter, or by using only the distance measure based on the power level difference minimization, the gain codebook is calculated. Search and select the optimal gain code.

In implementing the present invention, a dedicated processor for signal processing may be used to realize the hardware, or software may be realized in the form of a computer program.

[0016]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 shows an example of the functional configuration of the encoding method according to the present invention, in which parts corresponding to those in FIG. 3 are designated by the same reference numerals. Further, FIG. 2 shows an example of the functional configuration of the decoding method according to the present invention, and the parts corresponding to those in FIG. In order to explain the image of the present invention more clearly, the decoding method of FIG. 2 will be described first. Further, the present invention can be applied not only to audio signals but also to audio signals such as music signals, but in the following description, audio signals are used as a representative.

The audio signal reproduced by the decoding method according to the present invention is represented by the alternate long and short dash line in FIG.
4. The background noise stationary component generation unit 45 is composed of all or a part of the signal generated. Since the conventional decoding method shown in FIG. 4 has only the driving sound source vector generation unit 35 among the above three generation units, the naturalness of the reproduced sound when the speech signal with background noise is encoded and decoded. Was not enough. The reason why the naturalness is not enough is
This is because, as described above, the model of the driving sound source vector generation unit 35 is based on the model of the mechanism by which humans generate speech, and is not necessarily suitable for modeling speech with background noise. On the other hand, the present invention is based on a model in which a reproduction signal represented by a noise model is added to a reproduction signal represented by a voice generation model (a component by a driving sound source vector). The model is expressed by two components: a "background noise fluctuation component" whose characteristics fluctuate in a short time, and a "background noise stationary component" whose characteristics fluctuate with a slow period in time, or does not fluctuate.

First, the code transmitted or accumulated in the encoder is the driving excitation code b in the code receiver 31.
c, linear prediction parameter code ba, mode information bm
Is decomposed into. The mode information bm is information obtained by classifying the noise-added voice signal into sections such as vowel sections, consonant sections, silent sections, and background noise sections. However, since it is difficult to accurately divide the input speech signal having background noise into the above sections, the ambiguous frame may be unclassified.

The drive excitation vector generation unit 35 is composed of an adaptive codebook 36 and a fixed codebook 37 as in the conventional method, and each codebook corresponds to a periodic code and a fixed code in the received drive excitation code. Adaptive code vector c
a, a fixed code vector cr is output. These vectors ca and cr are multiplied by the gains ga and gr output from the gain codebook 41 corresponding to the gain code in the received drive excitation code, respectively, and then added to form the drive excitation vector c.

The vector output from the background noise variation component generation unit 44 corresponds to the component of the background noise in which the power and spectral characteristics vary in frame units. The white noise generator 51 outputs a white noise series vector s1 having a length of one frame. Gaussian random numbers are generally used as the white noise generation method, but uniform random numbers are used, or pseudo-methods such as storing random number sequences in a table in advance and cutting them out in frame length units are used. Good.

The driving sound source vector c output from the driving sound source vector generation unit 35 is input to the sound source power measuring unit 52, and the driving sound source vector c in the frame.
Of the white noise vector s1 based on the measured power Pc and the mode information bm.
And the weight w by which the driving sound source vector c is respectively multiplied
The weight creating unit 53 determines s1 and wc. White noise vector s1 and driving sound source vector c
Are the weights ws1 and wc, respectively.
After being multiplied by 4, 55, they are added by the adder 56 to be the input vector cs1 to the synthesis filter 33.

The weights ws1 and wc created by the weight creating unit 53 are the synthesis filter input cs.
The power of 1 is determined to be the same as the power of the driving sound source vector c (they do not need to be exactly the same, but they are perceived to have the same power). If the frame is a vowel section or if the section is unknown, w
s1 is 0 or a small value close to 0. This is because in the vowel section, the input speech is sufficiently well expressed by the driving sound source vector generation unit 35 based on the speech generation model, and it is not necessary to add the background noise fluctuation component. Also,
Even when the section classification is uncertain, it is safer not to add the background noise fluctuation component.

In FIG. 2, the output of the background noise variation component generation unit 44 is added to the driving sound source vector before the synthesis filter 33. However, when the synthesis filter 33 is a linear filter, the background noise variation component generation unit is used. The result obtained by passing the output of 44 through the synthesis filter and the one obtained by passing the driving sound source vector through the synthesis filter 33 (that is, even after the synthesis and the addition) are equivalent.

The output ys1 of the synthesizing filter 33 has the spectral envelope and the pitch component emphasized in the post-processing unit 34 as in the conventional method. However, since it becomes unnatural when emphasized except the voice section, only the vowel section of the mode information bm is emphasized to the same extent as in the conventional method, and the degree of emphasis is weakened as the vowel characteristic becomes lower. The output ys1 of the synthesis filter 33 is also sent to the average noise power measurement unit 61 of the background noise stationary component generation unit 45.
The average noise power measurement unit 61 measures the power of the output signal ys1 of the synthesis filter 33 only in a certain background noise section (including a silent section), and obtains the average value of the power over a sufficiently long time with respect to the frame length. calculate.
Here, the “certain background noise section” means that a frame in which a vowel section or a consonant section is suspected is excluded even in the background noise section. The switch 46 is controlled by the mode information bm so that the output of the synthesizing filter 33 is supplied to the average noise power measuring unit 61 only in the certain background noise section. Also, "a sufficiently long time for the frame length" means
It is considered that about 1 second to several tens of seconds is preferable. As a method of calculating the long-term average, the power of each frame may be stored in a buffer, and the average may be taken at regular intervals.
Let P (n) be the instantaneous power in the nth frame, Pave (n) be the average power in the nth frame, Pave (n-1) be the average power in the n-1th frame, and Pave (n ) = (1−α) Pave (n−1) + αP (n). The smaller the value of α, the longer the average. The calculated average noise power Py is sent to the weight creating unit 62,
The weight wu that determines the power of the background noise stationary component is calculated. The power of the signal output from the background noise stationary component generation unit 45 is determined to be approximately the same as the average noise power Py, but if it is determined to be slightly lower, it often sounds easier to hear. The power of the output u2 of the synthesis filter 63 is the average noise spectrum a.
Since it depends on u, when the weight creating unit 62 obtains the weight wu, the average noise power Py and the filter gain of the synthesis filter 63 are used together, or the power of the output u2 of the synthesis filter 63 is actually measured and Wu based on the value
You should ask.

Similar to the average noise power measuring section 61, the average noise spectrum measuring section 64 is sufficient for the frame length from the decoded linear prediction parameter aq only in the certain background noise section (including the silent section). Calculate the average value of the spectrum over a long time. Therefore, the switch 47 is controlled by the mode information bm so that the linear prediction parameter aq is supplied to the average noise spectrum measurement unit 64 only in the reliable background noise section. Generally, the average value of the spectrum is often averaged in the line spectrum pair (LSP) region, which is a type of linear prediction parameter, but may be averaged in the cepstrum and power spectrum regions. As a method of calculating the average, similar to the average of the power, the spectrum parameter for each frame may be stored in the buffer, and the average may be taken at a constant time, or a sequential update formula may be used. From the average noise spectrum measuring unit 64, a linear filter coefficient au corresponding to the average noise spectrum is output and becomes a coefficient of the synthesis filter 63. The method of calculating the linear filter coefficient from the line spectrum pair (LSP) is also described in "Digital Speech Processing" by Sadahiro Furui (Tokai University Press).

The white noise generator 65 is a white noise generator 5.
Similar to 1, outputs white noise s2 of 1 frame length,
The output white noise vector s2 is used in the synthesis filter 6
It is passed through 3 and becomes u2. The stationary sound generator 66 outputs a completely constant sound u3 that does not depend on the property of the input signal. The output of the background noise steady component generation unit 45 is obtained by adding the output signal u2 of the synthesis filter 63 and the signal obtained by multiplying the amplitude of the steady sound u3 by the weighting wu3 in the multiplication unit 67 in the addition unit 68, The weight wu created by the weight creating unit 62 is multiplied by the multiplying unit 69. Since the signal u3 output from the stationary sound generator 66 is a sound independent of the input signal, from the standpoint of encoding, decoding and reproducing the input signal, it may or may not be used. Good. However, in terms of human auditory sense, stationary background noise provides a sense of security, and even if the background noise does not always match the characteristics of the input signal, it often feels more natural.
Therefore, the stationary sound generating unit 66 generates a low-frequency sound such as a boon and a stationary white noise such as a sir, and adjusts the amplitude to a level relatively lower than the output level of the synthesis filter 63 and adds the noise. Then it becomes more natural background noise.

The output of the background noise stationary component generating section 45 is added to the output signal ys1e of the post-processing section 34 by the adding section 71 and becomes a reproduction signal output. The important thing to note here is
The level of the background noise stationary component is determined by the measurement result Py of the average noise power, regardless of the mode for each frame such as the vowel section, the consonant section, and the background noise section.
It is to be added at an almost constant level. This point is
The difference is that the background noise fluctuation component controls the added noise level for each frame depending on the mode information of the corresponding frame.

The output of the background noise stationary component generator 45 may be added between the synthesis filter 33 and the post-processor 34, but the post-processor 34 is a process for emphasizing the voice. When the degree of speech enhancement is large, the processing is easier and the reproduced speech is more natural when added after the post-processing unit 34 as shown in FIG. Next, the encoding method will be described with reference to FIG.

In order to obtain a natural reproduced voice by using the decoding method as shown in FIG. 2, the coding side must realize the following two points in addition to the conventional coding method. . The first point is to divide modes into frames, such as vowel sections, consonant sections, silent sections, and background noise sections, and send all or part of the mode information to the decoding side. The second point is background noise sections. In, the noise power information is sent to the decoding side.

In FIG. 1, the input audio signal x is also sent to the mode judging section 81. The mode determination unit 81 analyzes the input signal and outputs the mode information bm indicating the characteristic of the section.
Is output. As for the mode classification on the encoding side, the category "I don't know (unknown)" is also allowed in four of "vowel interval", "consonant interval", "silence interval", and "background noise interval". I will decide. The reason why the unknown is allowed is that when an audio signal on which background noise is superimposed is analyzed, there is always a frame with an intermediate property, so it is forced to
It is considered that it is not appropriate to classify the sound into one of the two intervals from the standpoint of reproducing a natural sound. However, if the 5 pieces of mode information including unknown are sent to the receiving side, the transmission information is wasted. Therefore, the unknown mode is used only by the encoding side and “unknown” is transmitted.
May be included in the vowel segment. When sending the above-mentioned four mode information, generally 2 bits are required, but if the first or not vowel is expressed by 1 bit, the information other than the vowel section requires less information than the vowel section. Since the input signal can be expressed, that is, there is a surplus in the used bits, the surplus bits can represent each consonant / silence / background noise section. Would require the use of one extra bit per frame as compared to the conventional method. 1 per frame
A bit is, for example, 5 if the frame length is 20 milliseconds.
This corresponds to 0 bit / sec, and a slight increase in the amount of information is sufficient with respect to the total amount of information of 4 kbit / sec.

The modes are vowel section / consonant section / silent section /
As a method of dividing into the background noise section, the power of the signal, the fluctuation of the power, the slope of the spectrum envelope, the pitch periodicity, etc. are obtained by analysis, and each is compared with a threshold value for judgment. Further, in the case of the background noise section, since the characteristics of the signal are various, it is preferable to consider the continuity of the modes or to use the relative threshold value in comparison with the power and characteristics of the past background noise section. . If clear interval classification is not possible even if the values obtained by analysis are compared with the respective thresholds,
The section is “unknown”.

In order to send noise power information to the decoding side in the second background noise interval, in the present invention, each of the distortion calculation section 16 and the power level difference calculation section 82 is searched for in the gain codebook 23. Search the codebook using the weighted sum of the outputs. In the CELP-based coding system, the power of the input signal is not necessarily reflected in the power of the reproduced sound, especially in the background noise section and the consonant section. This is the conventional CEL
In contrast to the fact that the P-type coding scheme codebook search is performed with the aim of reducing the waveform distortion in sample units, the coding model does not match the background noise or consonant generation process. to cause. Therefore, when the conventional method is used, in the background noise section and the consonant section, not only does the power of the reproduced signal not accurately represent the power of the input signal, but it is often unnatural and unstable. In the present invention, the decoding side calculates the power of the synthesis filter output ys1 and estimates the background noise level Py. Therefore, if the conventional codebook search method is used, the decoding side may estimate an incorrect noise level. Will end up. Therefore, on the encoding side in the present invention, the output y of the synthesizing filter 14 is sent to the same waveform distortion calculating unit 16 as in the conventional case to calculate the waveform distortion value d for each sample with the input signal, and the synthesizing filter output y is the power level difference calculation unit 82
Also, the difference between the power of the input signal and the power of the combined signal y is calculated. Generally, in CELP system coding, an optimum combination of an adaptive code vector ca, a fixed code vector cr, and gains ga and gr by which they are multiplied is searched, but in reality, it is enormous to search for these simultaneous optimum values. Since a large amount of calculation is required, the adaptive codebook 18 and the fixed codebook 19 are searched first to determine the optimum or suboptimal adaptive code vector ca and the fixed code vector cr, and then the gain codebook 23 is finally set. I often search.

In the present invention as well, although the above search order is used, the search for the adaptive codebook 18 and the fixed codebook 19 is performed by searching the codebook based on the minimization of the waveform distortion d as in the conventional case. When searching the codebook 23, the waveform distortion d
And the power level difference dp are used in combination. In the weight creation unit 83, the weights wd and wp by which the respective values of the waveform distortion value d and the power level difference dp are multiplied by the multiplication units 84 and 85 based on the mode information bm of the frame.
To create. For example, in the vowel section, only the waveform distortion value d is used, or a set of weights giving priority to the waveform distortion value d is used.
In the consonant section and the background noise section, only the power level difference dp is used or a set of weights giving priority to the power level difference dp is used. As a result, in the vowel section, a good quality sound with less waveform distortion as in the conventional case is obtained, but in the consonant section and the background noise section, the shape of the composite waveform y is highly similar to the shape of the input signal x, and the power is the input signal. A drive excitation code is selected for reproducing a sound that preserves the power of s. Since the "unknown" section has an intermediate property between the vowel section and the consonant or background noise section, it is preferable to use a weight that emphasizes the power level difference dp as the weight for searching the gain codebook 23. . The waveform distortion value d and the power level difference d weighted by the multiplication units 84 and 85, respectively.
p is added by the addition unit 86 and input to the codebook search control unit 25.

[0034]

In order to investigate the effects of the present invention, the present invention was realized in the form of simulation by a computer program, and subjective quality evaluation experiments were conducted using actual voice data. The frame length was 20 milliseconds and the frame was divided into two subframes. The bit rate was designed at 4 kilobits / second. As shown in FIG. 5, the detailed bit allocation per frame is a linear prediction parameter of 20 bits, an adaptive codebook of 13 bits, and a fixed codebook of 34 bits. The gain codebook is 12 bits in the vowel section and is in the sections other than vowels. It is set to 10 bits. The mode information has 1 bit in the vowel section and 3 bits in the section other than the vowel.

The voice data conforms to a general telephone characteristic called a modified IRS characteristic and has an SN ratio of 15d.
The vehicle noise-added speech of B and the office noise-added speech having an SN ratio of 30 dB were encoded and decoded using the conventional method and the present invention, respectively, and the reproduced sounds were actually heard and compared. The test subjects were 24 ordinary people, and the test method was very bad (-3), bad (-2), a little bad (-1), the same quality (0), comparing the quality of the original sound and the coded speech. It was rated on a 7-point scale: slightly good (+1), good (+2), and very good (+3).

The test results are shown in FIG. The lower the graph, the poorer the quality of the original sound. From the figure, it is shown that the quality when the present invention is used is greatly improved as compared with the quality obtained by the conventional method which does not use the present invention. Expressing the difference in sound properties between the conventional method and the present invention in words, the reproduced sound by the conventional method was very unnatural and unpleasant in the consonant section and the background noise section. The sound was reproduced with the atmosphere of automobile noise and office noise, and was of a natural quality that could be heard with peace of mind. Further, the present invention is superior in terms of a more natural reproduced sound in which voice and background noise are mixed not only in the background noise section but also in the vowel section.

When the present invention is applied to speech without background noise, "background noise" is classified as a silent section, and it is determined that the average level of "background noise" is 0. It goes without saying that it does not adversely affect. Even when an actual voice is input, the vowel section and the silent section (background noise section) have a sound quality equivalent to that of the conventional method, and the consonant section has a more natural sound quality according to the present invention.

[Brief description of drawings]

FIG. 1 is a block diagram showing a functional configuration example of an encoding device according to the present invention.

FIG. 2 is a block diagram showing a functional configuration example of a decoding device according to the present invention.

FIG. 3 is a diagram illustrating a conventional speech code-driven linear predictive coding (Code-E).
The block diagram which shows the function structure of a xcited Linear Prediction (CELP) apparatus.

FIG. 4 is a diagram illustrating conventional code-driven linear predictive coding (Code-E) for speech.
The block diagram which shows the function structure of the decoding apparatus corresponding to xcited Linear Prediction (CELP).

FIG. 5 is a diagram showing a bit combination in a simulation experiment.

FIG. 6 is a diagram showing a simulation experiment result.

─────────────────────────────────────────────────── ─── Continuation of front page (58) Fields surveyed (Int.Cl. ⁷ , DB name) G10L 19/12

Claims (10)

(57) [Claims] 1. A drive vector is generated by selecting a code vector for each frame from at least one of an adaptive codebook and a fixed codebook, and multiplying the selected code vector by a gain selected from a gain codebook. Then, in the acoustic signal decoding device that drives the synthesis filter with the driving sound source vector and outputs the acoustic signal, a means for receiving information indicating whether or not the frame is the background noise section, and the background noise section. A means for calculating a power level representing a long-term average value of the output signal power from the synthesis filter in a frame; and an average spectrum representing a long-term average of the spectrum from the filter coefficient of the synthesis filter in a frame which is a background noise section. And a second sum driven by white noise, the average spectrum being provided as filter coefficients. A filter, a means for amplitude-adjusting the signal generated by the second synthesis filter based on the calculated power level to generate a stationary component signal of background noise; Means for adding a signal to the output signal from the synthesis filter to output a reproduced sound signal regardless of whether it is a noise section, and whether the frame is a background noise section or a consonant section
For measuring the power of the driving sound source vector of the frame concerned,
Means, the measured power, and the background noise section or consonant section
The first and second weights based on information indicating whether or not
Of the background noise by multiplying the white noise by the first weight.
Means for generating a split signal, and means for multiplying the driving sound source vector by the second weight
And in the frame that is the background noise section or consonant section,
Driving source obtained by multiplying the variable component signal by the second weight
Means for adding to the vector, the power of the fluctuation component signal, and the second weight
The sum of the power of the driving sound source vector multiplied by
The first and second weights should be equal to the specified power.
The audio signal decoding device is characterized in that only the audio signal is determined . 2. A stationary component signal obtained by adding a stationary component signal from the means for generating a stationary sound having a constant spectral characteristic and a constant amplitude, and a stationary component signal generated by the means for generating a stationary component signal of the background noise. 2. The acoustic signal decoding device according to claim 1, further comprising: 3. A drive vector is generated by selecting a code vector on a frame-by-frame basis from at least one of an adaptive codebook and a fixed codebook, and multiplying the selected code vector by a gain selected from a gain codebook. Then, in the acoustic signal encoding device that compares the acoustic signal obtained by driving the synthesis filter with the driving sound source vector and the input acoustic signal, and selects the code vector and the gain, the input acoustic signal is analyzed. , The frame is background noise
And a means for determining whether it corresponds to a consonant section or a vowel section and outputting a code representing the characteristic of the determined section; and a power level difference between the output signal of the synthesis filter and the input acoustic signal. means for, upon said gain selection, the distance measure the power level difference in the frame of the background noise period and consonant segment, mother
Input signal of the output signal of the synthesis filter in the sound section
And a means for selecting a gain that minimizes the distance measure from the gain codebook , using the waveform distortion with respect to as the distance measure, and the acoustic signal encoding device. 4. In the vowel section, the waveform distortion is prioritized and
Within the scene noise section or consonant section, the power level difference is
Weighted sum is the distance between the preceding waveform distortion and power level difference
The acoustic signal encoding device according to claim 3 , wherein the acoustic signal encoding device is used as a scale . 5. A drive vector is generated by selecting a code vector on a frame-by-frame basis from at least one of an adaptive codebook and a fixed codebook, and multiplying the selected code vector by a gain selected from a gain codebook. Then, in the acoustic signal decoding method for driving the synthesis filter with the driving sound source vector to output the acoustic signal, the process of receiving information indicating whether the frame is the background noise interval and the background noise interval. Calculating a power level representing a long-term average value of the output signal power from the synthesizing filter in a frame; calculating an average spectrum representing a long-term average of spectra from filter coefficients of the synthesizing filter; Driving the second synthesis filter with white noise as a filter coefficient, and generating the second synthesis filter. A step of amplitude-adjusting the signal based on the calculated power level to generate a stationary component signal of background noise; and combining the stationary component signal regardless of whether the frame is a background noise section or not. The process of adding the output signal from the filter and outputting the acoustic signal, and whether the frame is in the background noise section or the consonant section
And the power of the driving sound source vector of the frame is measured.
Process, the measured power, the background noise section or the consonant section
The first and second weights based on information indicating whether or not
Of the background noise by multiplying the white noise with the first weight.
Generating a split signal and multiplying the driving source vector by the second weight
And the fluctuation component in the background noise section or the consonant section frame.
A signal to the drive source vector multiplied by a second weight
And the power of the fluctuation component signal is multiplied by the second weight.
The sum of the power of the driving source vector
The first and second weights to be equal to
An acoustic signal decoding method characterized by: 6. A step of generating a stationary sound having a constant spectral characteristic and a constant amplitude, and a step of adding the stationary component signal of the background noise and the stationary sound and outputting as a stationary component signal. The audio signal decoding method according to claim 5 , wherein 7. A code vector is selected from at least one of an adaptive codebook and a fixed codebook on a frame-by-frame basis, and a gain selected from each gain codebook is multiplied by the selected code vector to obtain a driving excitation vector. In the acoustic signal coding method, which generates and compares the acoustic signal obtained by driving the synthesis filter with the driving sound source vector and the input acoustic signal, and selects the code vector and the gain, the input acoustic signal is analyzed. The relevant frame is the background noise section
And a process of determining whether it corresponds to a consonant section or a vowel section and outputting a code representing the characteristic of the determined section, and calculating the power level difference between the output signal of the synthesis filter and the input acoustic signal. And the gain selection, the power level difference is used as a distance measure in the frames of the background noise section and the consonant section, and the vowel section is set.
In the waveform of the output signal of the synthesis filter with respect to the input signal
An acoustic signal encoding method comprising: a step of selecting a gain that minimizes the distance measure from a gain codebook using distortion as a distance measure. 8. In the vowel section, prioritizing the waveform distortion,
In the noise or consonant section, the power level difference is
Wherein a weighted sum of the previous waveform distortion and path Wareberu difference distance
The acoustic signal encoding method according to claim 7, which is used as a measure . 9. A computer-readable recording medium that stores a program for causing a computer to execute the acoustic signal decoding method according to claim 5 or 6 . 10. The method of claim 7 or computer readable recording medium an audio signal coding method for storing a program which Ru is executed by a computer as described in 8.