JP3199128B2 - Audio encoding method - Google Patents

Abstract

PURPOSE:To obtain an encoded voice of auditorily high quality with a small number of bits even where the periodicity of a voice is unsteady or even when the periodicity can not be reproduced sufficiently by a mode of an adaptive code book. CONSTITUTION:Fluctuations L1, L2, and L3 of the pitch period in a frame of the residual signal (a) generated by inversely filtering an input voice through a synthesizing filter are detected and the input voice is converted into a voice with the same pitch L4=(L2+L3)/2 as shown by (b) or a voice with the same pitch period L5=(L1+L2+L3) as shown by (c), and, the distortion of a synthesized voice at the time of the adaptive vector selection of the adaptive code book is calculated for the converted voice and an adaptive vector which minimizes the distortion is selected.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a digital encoding system for a speech signal sequence with a small amount of information by performing a predictive encoding for synthesizing a speech by driving a filter representing a spectral envelope characteristic of the speech with a sound source vector. The present invention relates to a highly efficient speech coding method.

[0002]

2. Description of the Related Art In order to efficiently use radio waves in digital mobile communication and to efficiently use a storage medium in a voice storage service or the like, a high-efficiency voice coding method is used. At present, as a method of encoding speech efficiently, an original speech is divided into sections called frames, which are arranged at regular intervals of about 5 to 50 ms, and the speech of one frame is subjected to a linear filter characteristic representing an envelope characteristic of a frequency spectrum. And a drive excitation signal for driving the filter are separated into two pieces of information, and a method of encoding each of them is proposed. In this method, as a method of encoding the drive excitation signal,
There is known a method of separating and encoding a drive excitation signal into a periodic component considered to correspond to a pitch period (fundamental frequency) of a voice and other components (non-periodic components). Code-Excited Linear Prediction (CELP) is an example of an encoding method of the driving excitation information. For details of the above technology, refer to the document MRSchroeder an
d BSAtal, "Code-Excited Linear Prediction (CELP):
High-Quality Speech at Very Low Bit Rates ", IEEE Pr
oc.ICASSP-85, pp.937-940,1985.

FIG. 8 shows a configuration example of the above-mentioned encoding method. For the original speech input to the input terminal 1, the linear prediction analysis unit 2 calculates a linear prediction parameter representing the frequency spectrum envelope characteristic of the original speech. The obtained linear prediction parameters are encoded by the linear prediction parameter encoding unit 3 and sent to the linear prediction parameter decoding unit 4. The linear prediction parameter decoding unit 4 reproduces a filter coefficient from the received code and sends it to the synthesis filter 11 and the distortion calculation unit 12. The details of the linear prediction analysis and examples of encoding of the linear prediction parameters are described in, for example, “Digital Speech Processing” by Sadahiro Furui (Tokai University Press). Here, the linear prediction analysis unit 2, the linear prediction parameter encoding unit 3, the linear prediction parameter decoding unit 4, and the synthesis filter 11 may be replaced with a nonlinear analysis and a nonlinear filter.

From the adaptive codebook 5, the immediately preceding past excitation vector (already quantized one to several frames of the excitation vector) stored in the buffer is converted to a length corresponding to a certain period. , And by repeating the cut vector until the length of the frame,
Time-series vector candidates corresponding to the periodic components of the voice are output.

[0005] From the noise codebook 6, candidates for a time-series code vector having a length of one frame corresponding to the non-periodic component of speech are output. These candidates are usually based on white Gaussian noise, and a predetermined number of candidate vectors according to the number of bits for encoding are stored independently of the input speech. Candidate of each time series vector outputted from the adaptive codebook 5 and the random codebook 6, in the multiplication section 8,9, the weight g _a created in the weight creating unit 7, _{respectively,} g _r is multiplied, adding section 10 Are added in
It becomes a candidate of the driving sound source vector.

The synthesis filter 11 is a linear filter that uses the output of the linear prediction parameter decoding unit 4 as a filter coefficient, and receives a driving excitation vector candidate output from the adding unit 10 as an input and outputs a reproduced voice candidate. Synthesis filter 1
The order of 1 or the order of the linear prediction analysis is generally 10 to 10.
The 16th order is often used. As described above, the synthesis filter 11 may be a non-linear filter.

FIG. 9 schematically shows the manner in which one playback audio candidate is created by the above procedure. Distortion calculator 12
Then, the distortion between the reproduced voice candidate output from the synthesis filter 11 and the input voice is calculated. The calculation of this distortion is
For example, it is often performed taking into account the coefficients of the synthesis filter 11 such as auditory weighting or the linear prediction coefficients that have not been quantized.

[0008] The codebook search control unit 13 generates a cyclic code that minimizes distortion between each reproduced speech candidate and the input speech.
A noise code and a weight code are selected, and a driving excitation vector in the frame is determined. Codebook search control unit 13
The periodic code, the noise code, the weight code, and the linear prediction parameter code output from the linear prediction parameter encoding unit 3 determined in are sent to the code transmission unit 14 and stored according to the form of use. Or sent to the recipient.

FIG. 10 shows a configuration example of a decoding method. The decoding method is different from the encoding method in that a speech is reproduced from the received code without searching the codebook, but has basically the same configuration as the encoding method. The code receiving unit 21
A linear prediction parameter code from a transmission line or a storage medium,
Receives periodic code, noise code and weight code. The linear prediction parameter code is sent to the linear prediction parameter decoding unit 22 and is converted into a filter coefficient.

The adaptive codebook 23 derives the immediately preceding past excitation code from the driving excitation vector of the previous frame stored in the buffer (the already-decoded driving excitation vector of one to several frames immediately before) and the received periodic code. The driving excitation vector is cut out at a length corresponding to the cycle specified by the cycle code, and the cut-out vector is repeated until the length of the frame is reached, thereby outputting a time-series vector corresponding to the sound cycle component.

The noise codebook 24 outputs a time-series vector specified by the received noise code from time-series vectors of a length corresponding to one frame stored in a predetermined number. The weight creation unit 25 outputs a value of a weight to be multiplied by a code vector output from the adaptive codebook 23 and the noise codebook 24 from the received weight code.

Each of the above code vectors is multiplied by a multiplication unit 2
At 6, 27, the weight value is multiplied. The two vectors multiplied by the weights are added in the adding unit 28 to become a driving sound source vector. The synthesis filter 29 is a linear filter using the output of the linear prediction parameter decoding unit 22 as a filter coefficient, receives the driving excitation vector from the addition unit 28, and reproduces a sound. As described in the section of the encoding method, a non-linear filter may be used as the synthesis filter 29 in some cases.

The reproduced sound is output from a terminal 30.

[0014]

In such an encoding method, an adaptive codebook plays a large role. In a portion such as a voiced portion in which the periodic component of speech is large, the value of the weight g _a multiplied by the adaptive code vector output from the adaptive codebook is generally coded to a large value (close to 1). On the other hand, in places where the voice has few periodic components such as unvoiced parts, the weight g
The value of _a is encoded into a small (close to zero) value. However, even in a portion where the speech has periodicity, for example, when the periodicity is non-stationary or when the adaptive codebook model cannot sufficiently represent the periodicity, the distortion calculated by the distortion calculator 12 is Codebook search control unit 13 to minimize
When the value of weight code = weight g _a is determined by the following equation, the value of g _a is coded small, and the quality of the reproduced voice is significantly reduced. An object of the present invention is to improve the coding part by the adaptive codebook even in a case where the periodicity of the voice is non-stationary or in the case where the periodicity cannot be sufficiently expressed by the model of the adaptive codebook. It is an object of the present invention to provide a method for encoding audio with high acoustic quality.

[0015]

According to the present invention, in a case where the periodicity of speech is non-stationary or when the adaptive codebook model cannot sufficiently express the periodicity, the input speech is matched with the adaptive codebook model. Then, by encoding the periodic code of the adaptive codebook and the weight code = weight g _a , the value of the weight g _a is greatly encoded to realize highly efficient encoding. In order to minimize the deterioration of auditory quality, by performing the deformation itself within a range that has almost no auditory effect, or by performing processing to compensate for the deterioration caused by the deformation in the decoding process, A low bit, perceptually high quality speech encoding method is realized.

[0016]

FIG. 1 shows an example of the configuration of an encoding method according to the present invention, and portions corresponding to those in FIG. According to the present invention, the original audio input from the terminal 1
In 1, the signal is transformed into a signal that can be more easily represented by the model of the adaptive codebook 5 and sent to the distortion calculator 12. This preprocessing converts the continuous or fluctuating voice components contained in the input voice, which are defined in units of a pitch period and a waveform having a length corresponding to the pitch period, into a steady signal within a frame. The specific example will be described later. The distortion calculator 12 calculates the distortion between the pre-processed voice and each of the reproduced voice candidates, and calculates the codebook search controller 1.
In step 3, each code that minimizes distortion is selected.

Depending on the preprocessing method, the preprocessing may be performed using the driving sound source vector of the previous frame. If the deformed pre-processed speech has almost no auditory deterioration, there is no need to transmit a code for restoration, and the conventional decoding method shown in FIG. be able to. On the other hand, in order to greatly improve the efficiency of encoding, preprocessing that may cause auditory deterioration may be required. In this case, the restoration code corresponding to the processing in the pre-processing unit 31 is transmitted from the code transmission unit 14, and the synthesis filter 29 is decoded by the decoding method shown in FIG. The restoration processing unit 32 performs restoration processing to compensate for the deterioration in quality due to deformation using the restoration code. However, if the deterioration due to the pre-processing can be compensated for by the predetermined processing without using the repair code, there is no need to send the repair code.

In the decoding method shown in FIG. 2, the restoration processing section 32 is inserted after the synthesis filter 29, but may be inserted between the addition section 28 and the synthesis filter 29. In the encoding method of FIG. 1, the original speech is preprocessed and used for distortion calculation. However, the reference speech input to the distortion calculator 12 is the original speech and the driving Equivalent processing can be performed by modifying the sound source vector candidates.

The portions shown by broken lines in FIGS. 1 and 2 indicate that they are used as needed. In code driven linear predictive coding (CELP), the reference (input) speech vector is x, the adaptive code vector is e, the weight of the adaptive code vector is g _a , the noise code vector is c, and the weight of the noise code vector is g. _r , a matrix having the impulse response of the synthesis filter 11 as an element

[0020]

(Equation 1) Then, the distortion E is generally defined as follows. _{E = ‖x-g a He-} g r Hc‖ 2 However, terms such as perceptual weighting is omitted. Even when the auditory weighting is considered, the expression can be expressed in the same form as the above equation by only slightly changing x and H. In the above equation, E
Search for g _a, e, g _r , and c so that is minimized.
However, in general, in order to search with a realistic amount of computation, g _a , e at which E is minimized as g _r Hc = 0 is determined first, and then g _a , e is determined.
In many cases, g _r and c are determined while fixing the values (or vectors) of g _a and e. The optimum value of g _a according to the above ^{_{method, g a = (x t He}} ) / ‖He‖ 2 strain, E = ‖x‖ ² - is given by ^{(x t He) 2 / ‖He‖} 2. Here, e is an unknown vector, and an optimal vector is searched for by an adaptive codebook model. At this time, if the value of g _a is small despite the fact that the portion is a voiced portion, the quality of the reproduced sound is generally poor. Therefore, if g _a can be increased by the pre-processing for deforming the vector x, the efficiency of encoding is improved and the quality of reproduced sound is improved.

Next, a specific example of the pre-processing unit 31 in FIG.
Examples will be described. Example of preprocessing for pitch period first
Is schematically shown in FIG. (A) is a synthetic filter for input speech
This is a residual signal that has been subjected to the inverse filter of the data 11. Audio is non
Since the signal is a steady signal, the pitch period is set for each pitch waveform.
Accurately, the pitch period is within the same frame
L_Two, L_ThreeAnd may have changed slightly. This is one
In addition to what is commonly called "fluctuations",
May change gradually due to the pitch change
You. Such a change in the pitch period is caused by the
As Dell cannot express it,_aIs the value of
Become smaller. Therefore, as shown in FIG.
Period equal intervals L_Four= (L_Two+ L_Three) / 2. is there
Or, as shown in FIG._Five= (L₁+ L_Two+ L
_Three) / 3. In CELP, the adaptive code
Lag (repeat unit) of number book is twice the actual pitch period
Or 3 times, so as shown in FIG.
And L₆= L ₀And L₇= (L₁+ L_Two+ L_Three−
L₀) / 2 when repeated with the adaptive codebook lag
Interval conversion may be performed so as to be optimal. (D)
The case of twice the pitch period is shown, but the case of three times and four times
Can be easily extended.

In these processes, as shown in FIG. 4A, the original speech is analyzed by a pitch cycle analysis section 33 to analyze the correct pitch cycle for each pitch waveform, and a pitch cycle conversion section 34 performs a speech waveform deformation process. I do. Further, FIG. 3 shows a state in which both the previous frame and the current frame are processed using the residual, but the past frame may be processed using the driving sound source signal of the previous frame. In this case, the pitch interval analysis unit 33 performs analysis using the driving sound source signal before the previous frame. In the above processing, when the conversion amount of the pitch period is small, there is almost no auditory deterioration due to the processing, so that there is no major problem even if the pre-compensation processing is not performed in the decoding process. However, in the course of decoding, it is even better to perform processing so that the pitch period again changes smoothly.

FIG. 5 shows an example of a pitch period analysis method.
FIG. 5A shows a residual signal obtained by applying an inverse filter of the synthesis filter 11 to the input voice. First, the maximum point in the residual of the current frame is found, and a one-pitch-length residual waveform is cut out centering on the maximum point. This cut-out waveform is shown in FIG.
(B). The one pitch length at this time may use an average pitch period in a frame calculated using, for example, a deformation correlation method or the like. In addition, the length to be cut out does not necessarily have to be one pitch length, and may be any length. Generally, the length is often about 70 to 80% of one pitch length.

Next, the extracted waveform is used as a matched filter coefficient while moving on the residual waveform as shown in FIG.
A correlation value series as shown in is calculated. At this time, the correlation value is 1 at the time of the maximum point at which the waveform is cut out. A reference position for determining an accurate pitch interval is determined by detecting a peak of the correlation value. The determined reference position is shown in FIG. The pitch of the pitch reference position determined in this way is defined as a pitch cycle (interval), and the reference position is used in subsequent deformation processing. In the above process, the residual shown in FIG. 5A may be replaced with a driving sound source signal that has already been quantized for a past frame. The residual of the current frame may be calculated by replacing the internal state of the inverse filter with the already-quantized drive excitation signal.

FIG. 6 shows an example of a method for converting the pitch period (interval). FIG. 6A shows the residual, and FIG. 6B shows the pitch reference position obtained by the above method. The conversion of the pitch interval is performed so that the interval between the reference positions becomes the specified value.
The pitch reference position shown in FIG. 6B is moved as shown in FIG. First, at each reference position in FIG. 6B, a window function having a weight of 1 as shown in FIG. 6D and a weight of 0 at an adjacent reference position is created. Cut out over the residual signal. Next, as shown in FIG. 6E, the cut out waveform is shown in FIG.
Is shifted in accordance with the movement of the reference position, and the shifted waveforms are superimposed again to generate a converted pitch interval residual. At this time, when performing the transformation process across frames, the voice is pre-read and the residual diagram
(A) must be stored in the buffer. Although a triangular window is used as the window function in FIG. 6, a Hanning window or the like may be used when emphasis is placed on saving a waveform near the reference position.

Although the above embodiment of the pitch interval conversion is based on the method of superimposing the waveforms, as shown in FIG. 7, the pitch interval can be converted by expanding and contracting the waveform from the reference position to the reference position. Good. As the simplest method, a method may be adopted in which a waveform between reference positions is deleted when the pitch is shortened, and 0 is inserted when the pitch is lengthened. The processed voice is created by passing the transformed residual through a synthesis filter.

Next, FIGS. 3E and 3F schematically show an example of preprocessing an amplitude (pitch gain) as a speech component defined in units of a waveform having a length corresponding to the pitch period. FIG. 3E shows a residual signal obtained by applying an inverse filter of the synthesis filter 11 to the input voice. Since voice is an unsteady signal, if the correlation (pitch gain) of a time-delayed waveform corresponding to the pitch period is accurately measured for each pitch waveform, the correlation value will change slightly even within the same frame. There is. That is, the peak values are slightly different. Even in such a case, it cannot be sufficiently expressed by the adaptive codebook model. Therefore, as shown in FIG. 3 (f), if the pitch gain is deformed so as to be constant within the frame and the same peak (amplitude) is obtained, the prediction accuracy by the adaptive codebook increases,
As a result, the value of the weight multiplied by the adaptive code vector can be increased.

In this process, as shown in FIG. 4B, the pitch gain analysis unit 35 accurately analyzes the pitch gain in the frame, and the pitch gain conversion unit 36 performs a deformation process. Also in this process, when the pitch excitation analysis unit 35 uses the driving sound source vector of the previous frame at the time of analysis, it is possible to perform an optimum process according to the quantization level. The pitch gain is converted by a method of multiplying each pitch waveform in the residual region by a certain value. In the case of this processing, since the quality may be slightly deteriorated due to the pre-processing, if necessary, in the decoding process, compensation processing is performed so that the pitch gain changes smoothly from the information of the preceding and following frames. It is desirable. However, it is considered that there is little need to allocate bits to the information for restoration and to particularly transmit the information as the restoration code.

Next, FIG. 3 shows an example in which the phase characteristic of the frequency spectrum is preprocessed as a speech component defined in units of a waveform having a length corresponding to the pitch period.
(G) and (h). FIG. 3 (g) shows that 1
FIG. 9 schematically shows a waveform obtained by cutting out a pitch waveform. The residual waveform generally has a white power spectrum, but non-stationarily changes in phase characteristics. For example, in the waveform of FIG. 3 (g), only the phase can be changed as shown in FIG. 3 (h) while maintaining the power spectrum (the state is schematically shown). Not only when the phase is changed in the frame, but even if the phase is different from the already-quantized drive excitation signal of the previous frame, it cannot be sufficiently expressed by the adaptive codebook model, so that the quality deteriorates.

Therefore, as shown in FIG.
In 7, a transformation process is performed in the phase converter 38 while analyzing the phase of the original voice for each pitch. In addition, when the phase analysis unit 37 analyzes the phase of the drive sound source signal of the previous frame together, and performs an optimal process such as adjusting to the phase of the previous frame, the effect is great. In FIG. 3H, the waveform in FIG. 3G is transformed into a waveform in which the pulse becomes steep, but in the modification of the present invention, the phase change is arbitrary. Since human auditory characteristics are insensitive to changes in phase information, extreme quality deterioration rarely occurs in this processing. However, in general, when the phase becomes monotonous, the reproduced sound becomes a buzzer sound, and tends to be a so-called “synthetic sound that lacks naturalness”. Therefore, it is considered that there is no particular problem even if the compensation processing is not performed if there are no spare bits for transmission or storage, but a more natural reproduction sound can be obtained by allocating some bits to the information for restoration. .

The above three pre-processings may be performed all at the same time or may be used in combination. In order to satisfy the adaptive codebook model more, the adaptive codebook repeats the past excitation vector in a certain cycle. It is good to transform to. Also in this case, the above-described preprocessing methods can be used together. In the case of this processing, since the input (reference) sound has a very monotonous waveform, it is considered that the so-called "synthetic sound and lack of naturalness" is effective in an application field in which the amount of information is prioritized. Can be However, when a certain degree of naturalness is required, it is necessary to encode the change between the processed voice and the original voice with an appropriate number of bits and transmit the resulting code as a repair code.

In addition, as a direct pre-process for increasing the value of g _a , the input x may be transformed using a dynamic programming technique. The present invention is applicable not only to CELP but also to all coding schemes based on an adaptive codebook model. Further, the present invention is also applied to a coding method of a type that expresses the periodicity of speech by a filter called a pitch filter having a pitch period as a tap position.

[0033]

As described above, according to the present invention, the expression of speech is facilitated by the adaptive codebook model, and as a result, the weight value (quantized value) multiplied by the adaptive code vector is increased. And distortion due to encoding can be further reduced. On the other hand, by performing pre-processing such that there is no or little information to be transmitted or stored to repair the deformation process, a higher quality code can be obtained without increasing the amount of information to be transmitted or stored. Can be realized.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an encoding method according to an embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of a decoding method according to the embodiment of the present invention.

FIG. 3 is a waveform chart for explaining various examples of preprocessing in the present invention.

FIG. 4 is a block diagram showing various configuration examples of a preprocessing unit 31.

FIG. 5 is a diagram schematically showing how a pitch reference position is determined in one embodiment for analyzing pitch intervals.

FIG. 6 is a diagram schematically showing an embodiment for converting a pitch interval.

FIG. 7 is a diagram schematically showing an example of a method of deforming a waveform from a reference position to a reference position in an embodiment for converting a pitch interval.

FIG. 8 is a block diagram illustrating a general configuration example of an encoding method of a code-driven linear prediction encoding method.

FIG. 9 is a diagram schematically showing a state in which a reproduced speech candidate is created from an adaptive codebook and a noise codebook in the code-driven linear prediction encoding method.

FIG. 10 is a block diagram illustrating a general configuration example of a decoding method of a code-driven linear prediction encoding method.

Continuation of front page (56) References JP-A-59-61891 (JP, A) JP-A-6-27996 (JP, A) JP-A-3-119398 (JP, A) JP-A-4-344699 (JP) , A) JP-A-61-150000 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 19/08 G10L 19/12

Claims (1)

(57) [Claims] 1. A filter comprising: a time-series vector extracted from an adaptive codebook, which is obtained by repeating a past excitation vector at a cycle corresponding to a pitch; and a time-series vector extracted from a noise codebook. A speech encoding method that encodes input speech by driving and reproducing the speech, wherein the pitch corresponding to the pitch cycle and the pitch cycle, which are included in the input speech, are continuously or fluctuating. A speech component defined as a unit of a waveform having a length, which is transformed into a stationary signal in a frame, and re-inputted as speech for encoding. .