JP3207466B2 - Coding method of pitch synthesis filter - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an encoding method of a pitch synthesis filter applied to CELP speech encoding.

[0002]

2. Description of the Related Art Conventionally, as a coding method capable of coding a speech signal at a low rate of about 8 kbit / sec and high quality, CELP (Cod
e Excited Linear Predicion) method is known. CE
For details on the LP method, see "A Class of Analysis-b" by P. Kroon et al.
y-Synthesis Coding at RatesBetween 4.8 and 16kbits
/ s ", IEEE SAC-6, pp. 353-363, February 1988.

[0003] The CELP system is classified into one of voice coding systems for generating a synthesized voice signal by using a driving signal, a synthesis filter, and a pitch synthesis filter. The feature of the CELP system is that a target vector generated from the voice signal is generated. An optimal code vector is to be selected from the drive signal codebook so that the waveform distortion between the signal and the composite vector is reduced.
FIG. 4 is a block diagram showing a coding part of pitch gain performed on the code side of the conventional CELP system and a part other than the coding part.

[0004] In the figure, reference numeral 212 denotes an input terminal, which is an input signal (for example, an audio signal) supplied to the input terminal 212.
Are sent to the filter coefficient calculation unit 201 and the target vector calculation unit 211, respectively.

A filter coefficient calculating section 201 performs linear prediction analysis of an input signal to obtain filter coefficients of a synthesis filter representing vocal tract characteristics, encodes them to obtain an impulse response, and converts the filter coefficient information into a target vector calculating section 211. , To the influence signal calculation unit 213 and the combined vector calculation unit 206, and further output from the output terminal 220. In addition, the target vector calculation unit 211 weights the input signal with a weight filter determined based on the number based on the filter coefficient of the filter coefficient calculation unit 201, and outputs the current signal based on the past encoding of the influence signal calculation unit 213 described later. An influence signal on the input signal is subtracted from the weighted input signal, and a target vector X used for vector quantization of the drive signal is calculated.

[0006] On the other hand, reference numeral 203 denotes a code book. The code book 203 sequentially takes in past drive signals generated by the drive signal generation unit 214,
By discarding unnecessary drive signals in 03, the contents can be updated. Here, the codebook 203 is CE
In the case of the LP method, a plurality of codes having different properties are often used, and an adaptive codebook known as the most efficient codebook is used here. This adaptive codebook uses a drive signal of voice coded in the past as a codebook. Since a drive signal in a section having strong periodicity (pitch) in human voice repeats a waveform similar to each pitch cycle. Thus, efficient encoding of the drive signal is enabled. The code vector of the adaptive code book is assigned a code number based on the pitch cycle, and a drive signal in the past from the current sample point by the pitch cycle is used as the code vector.

[0007] The output of the codebook 203 is sent to a combined vector calculation unit 206, which uses the code vector based on the impulse response of the filter coefficient calculation unit 201 and the pitch period T from the codebook 203. A convolution calculation is performed to calculate a composite vector _DT .

Furthermore, the synthetic vector D _T of the synthesis vector calculating unit 206 and input to the pitch synthesis unit 208 generates a pitch synthesis vector U _T subjected to pitch synthesis process giving a pitch period T in the search thereto. The pitch synthesis processing in this case can be performed by the following equation. U _T (n) = D _T (n) + U _T (nT) ... (1)

In the above equation, the processing of a pitch synthesis filter having a pitch period T and a pitch gain of 1 is represented. Such a pitch synthesis filter becomes 1 / (1-z- ^T ) in the z ^- conversion domain, and the input waveform Has only the effect of simply repeating with a period T.

Accordingly, in the conventional CELP scheme, the optimal gain should multiplied to U _T instead of the 1 pitch gain γ
_{Although T} is obtained and encoded, this method uses a pitch synthesis filter in which drive signals of exactly the same shape repeat within the update interval of the pitch period as shown in Expression 1, so that the same power is used. Although a repetitive voice signal can be expressed relatively efficiently, the voice cannot be accurately represented in a section where the power of the drive signal suddenly increases or a section where the power decreases at the end of a word.

[0011] Then, the resultant vector U _T given pitch period T in pitch synthesis unit 208, the code search unit 2
15, a composite vector U given a pitch period T
_A code search (pitch cycle search) is performed using _T and the target vector X. The code search here is performed by combining the combined vector γ _T U and the target vector X when the optimal gain γ _T is given.
And the distortion X ^t X- (X ^t U _T ) ² / (U _T ^t U _T ), that is, the value (X ^t U _T ) ² / (U _T ^t This is performed by searching for a code (pitch cycle) that maximizes U _T ). Then, the optimum gain value is provided to the gain determination unit 207, and the optimum code information is provided to the drive signal generation unit 207 and output to the output terminal 218.

In the gain determining section 207, the code searching section 2
The optimal gain γ _{T opt} = (X ^t U _Topt ) / (U
_Topt ^t U _Topt ) as the target value and the gain table 205
Is used to encode the gain of the code vector, and this is output to the output terminal 219. Also, the drive signal generation unit 214
Then, using the optimum code information of the code search unit 215 and the coded gain of the gain determination unit 207, the optimum code vector is multiplied by the gain to generate a speech drive signal for the current section. This drive signal is given to the influence signal calculation unit 213, and the influence signal calculation unit 2
At 13, an influence signal on the next section due to vector quantization of the audio drive signal of the current section is calculated.

According to the CELP system,
The synthesized voice signal generated by the selected drive signal code, the synthesis filter, and the pitch synthesis filter has been confirmed by the code side to have low distortion, so that the decoding side has a good bit rate of about 8 kbit / sec. It is possible to generate a complex synthesized voice.

By the way, in the conventional encoding at a bit rate of about 8 kbit / sec, adaptive codebook search and pitch synthesis are performed for each short section (subframe) of about 5 msec. When such a short subframe is used, the change in the gain and pitch cycle of the actual audio drive signal is reduced, so that the drive signal can be expressed relatively accurately.

On the other hand, recently, it has been considered to perform speech coding at a lower bit rate of about 4 kbit / sec. However, if such a low bit rate is used, the number of usable bits is reduced. Due to the restrictions, it is necessary to increase the length of the subframe to increase the pitch cycle and the pitch gain update cycle. But,
In this case, the gain and pitch cycle of the drive signal for each sub-frame change significantly, and the synthesized voice does not change smoothly, resulting in a deterioration in the quality of the synthesized voice.

[0016]

As described above, in the conventional speech coding, when the bit rate becomes as low as about 4 kbit / sec, the pitch cycle and the pitch gain update cycle (sub Frame), which makes it impossible to accurately represent drive signal information that changes within a sub-frame, especially pitch gain information, and cannot produce high-quality synthesized speech.

The present invention has been made in view of the above circumstances, and provides an encoding method of a pitch synthesis filter capable of generating a high-quality synthesized voice by smoothly changing the power of the synthesized voice even at a low bit rate of about 4 kbit / sec. The purpose is to provide.

[0018]

In order to achieve the above-mentioned object, a coding method of a pitch synthesis filter according to the present invention performs a pitch cycle search using an adaptive codebook, and specifies a pitch cycle by the searched pitch cycle. Speech encoding means for outputting a code vector of the adaptive codebook,
Synthetic vector generating means for generating a synthetic vector using a code vector output from a voice coding means, a recursive pitch synthesis filter using a pitch period obtained by an adaptive codebook search in the coding means and a pitch gain of a gain table as parameters And a means for performing a recursive pitch synthesis process on the synthesized vector of the synthesized vector generating means to generate a synthesized voice, and a means for selecting a pitch gain of the gain table based on a distortion amount of the synthesized voice. I have.

[0019]

As a result, according to the present invention, a recursive pitch synthesis filter is generated by using the pitch cycle determined by the closed loop pitch cycle search and the pitch gain candidate of the gain table in the CELP coding, and this pitch synthesis filter is used. By selecting the optimal pitch gain from the gain table based on the distortion amount of the synthesized voice signal, a synthesized voice having a smooth power change corresponding to a sudden increase / decrease of the power of the drive signal can be obtained. become. In other words, by using a recursive pitch synthesis filter using a high-precision pitch cycle obtained by CELP encoding, it is possible to provide a high-quality synthesized speech with a smooth power change and an accurate pitch cycle even at a low rate. Become like

[0020]

An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing a principle configuration of a CELP-type speech coding system to which the coding system of the pitch synthesis filter according to the embodiment is applied. In the figure, 112
Is an input terminal, and an input signal (for example, a voice signal) provided to the input terminal 112 is input to the CELP encoding unit 100.

As the CELP encoding unit 100, the one of the CELP system described with reference numeral 200 in FIG. 4 is applied. In addition, in the case of such a CELP system, a codebook such as a noise codebook is often used in addition to the adaptive codebook in order to encode the audio drive signal, but here, only the adaptive codebook is used. Is used.

Then, the synthesis filter coefficient extracted from the input speech signal in the CELP encoding section 100 is provided to the synthesis section 140 and output from the output terminal 120 at the same time. The information of the searched pitch cycle is
The signal is supplied to the delay unit 166 and output from the output terminal 118 at the same time. Further, the code vector from the adaptive code book specified by the searched pitch period Topt is provided to the synthesizing unit 140, and the target vector obtained by weighting the input signal is provided to the gain evaluating unit 170.

The synthesizing unit 140 _{obtains a} synthesized vector D _Topt based on the code vector _supplied from the CELP encoding unit 100 and the coefficients of the synthesis filter, and _{supplies the} synthesized vector D _Topt to the pitch synthesizing unit 160.

The pitch synthesizer 160 includes an adder 162, a delay 166, and a multiplier 164.
The gain vector 15
Pitch gain candidate value b from 0 and CELP encoding section 100
_Is processed by the recursive pitch synthesis filter P (z) shown in the following equation using the pitch period T _opt of
Generate Y. P (z) = 1 / (1- bz ^-Topt )… (2)

The pitch synthesis filter used here is a highly accurate pitch cycle T ^opt determined by a closed loop pitch cycle search in CELP coding. And pitch gain candidates in the gain table
b), and when b> 1, the amplitude of the input signal is smoothly increased by a factor of b for each pitch period T _opt , and b <
When it is 1, the input signal operates so as to smoothly decrease the amplitude of the input signal by a factor of b for each pitch period _Topt .

The operation of the pitch synthesizing unit 160 will be described in more detail.
_Topt (n) and the pitch gain candidate value b from the gain table 150 to the past value for the pitch cycle of the pitch synthesis vector.
Is added to the value bY (nT _opt ), and the pitch composite vector Y
(n) is output. Further, the delay unit 166 stores the past pitch synthesized vector, and outputs the value of the pitch synthesized vector in the past by the pitch period based on the pitch period T _opt from the CELP encoding unit 100. Then, in the multiplication unit 164,
A value obtained by multiplying this value by the pitch gain candidate value b from the gain table 150 is output to the adding unit 162. Such processing is performed, for example, according to the following equation. Y (n) = D _Topt (n) + bY (nT _opt ) (3)

Then, the gain evaluator 170 evaluates the amount of distortion between the target vector X and the pitch composite vector Y of the CELP encoder 100, and obtains a pitch gain b _opt that minimizes the amount of distortion.
Is searched from the gain table 150, information on the optimal pitch gain is provided to the CELP encoding unit 100, and is output from the output terminal 119.

Therefore, if the coding method of the pitch synthesizing filter as described above is applied, unlike a conventional pitch synthesizing filter in which drive signals of exactly the same shape repeat within a subframe, a recursive pitch synthesizing filter is used. Since it is possible to cope with a sudden increase / decrease of the power of the signal and to search for the gain in the gain table only for the determined pitch period, the increase in the amount of calculation required for the gain search can be negligibly small. There is also an advantage that the number of coded bits does not increase at all. This makes it possible to obtain high-quality synthesized speech having a smooth power change even in low-bit-rate speech encoding having a large subframe length.

Incidentally, in order to confirm the effect of the present invention,
A comparison experiment of coding was performed using actual speech signals, and the result of FIG. 3 was confirmed. Here, FIG. 3 (a) is a waveform diagram (60 msec) of an actual input voice signal. In the case of using a conventional pitch synthesizing filter in which drive signals of exactly the same shape are repeated within the update interval of the pitch cycle, FIG. Since the waveform of the synthesized voice signal does not change smoothly because of the waveform of the synthesized voice signal shown in (b), the synthesized voice signal does not resemble the power change of the input voice signal waveform. By using the synthesis filter, the power of the synthesized voice becomes a waveform of the synthesized voice signal that smoothly changes, as shown in FIG. 3C, and is closer to the power change of the input voice signal waveform. Next, a second embodiment of the present invention will be described with reference to FIG. In this case, in FIG. 2, portions denoted by the same reference numerals as those in FIG. 1 have the same functions as those described above, and thus description thereof will be omitted. In the figure, an input signal (for example, an audio signal) is input from an input terminal 112 to a filter coefficient calculator 101 and a target vector calculator 111.

The filter coefficient calculation unit 101 predicts and analyzes the input signal in units of a frame length (usually about 20 ms) to obtain a filter coefficient of a synthesis filter, encodes the filter coefficient, and outputs coefficient information to an output terminal 120. The processing below this is performed for each subframe, with the time length represented by the sample of the number of dimensions of the vector as one small section (subframe).

The target vector calculation section 111 weights the input signal with a weighting filter determined based on the filter coefficient of the filter coefficient calculation section 101, and adds the weighted input signal to the current input signal obtained by past coding of the influence signal calculation section 113 described later. Is subtracted from the weighted input signal to calculate a target vector X used for vector quantization of the drive signal.

On the other hand, the code book 103 is composed of an adaptive code book, and includes a drive signal generator 114 described later.
The content is updated by taking in the past drive signal V (n) generated in (1) and discarding unnecessary drive signals in the codebook 103. In this case, the code vector generated from the code book 103 is such that the code corresponds to the pitch period T = I (I is an integer) and the code vector V _I (n) (n = 0, 1,...) , L) is defined by the following equation. (Here, L represents the dimension of the vector.) V _I (n) = V (nI) (n = 0,1, ..., min (I-1, L-1)) ... (4 The code vector of the code book 103 is sent to the composite vector calculation unit 106.

In the composite vector calculation unit 106, the filter coefficients of the filter coefficient calculation unit 101 and the codebook 103
The code vector based on the search pitch period of calculating the combined vector D _I.

[0035] Then, provide a synthetic vector D _I of the synthetic vector calculation unit 106 to the pitch synthesis section 108. The pitch synthesizing unit 108 synthesizes the synthesized vector with a pitch using a pitch synthesizing filter having a pitch period I and a pitch gain of 1. Such processing is performed according to the following equation. U _I (n) = D _I (n) + U _I (nI) (5) When the pitch period I to be searched is equal to or greater than the number of dimensions L of the vector, the processing of the pitch synthesis unit 108 can be omitted.

The synthesized vector U of the pitch synthesizer 108
_IIs given to the code search unit 115. Code search section
At 115, the synthesized vector U of the pitch synthesis unit 108_IWhen
A code search is performed using the target vector X. here
Search for the value (X^t  U_I)^Two / (U_I ^t  U_I) Is the largest
This is done by searching for the code And optimal
Code information is output from the output terminal 118.
Is sent to the codebook 103. On the other hand, the combining unit 14
0 means that the optimal code vector and code
A synthesis filter based on the synthesis filter coefficient of the
Couture D_ToptAsk for.

Then, the synthesized vector D _Topt is sent to the pitch synthesizer 160. The pitch synthesizing section 160 receives the synthesized vector from the synthesizing section 140 as an input, and uses the pitch gain candidate value b from the gain table 150 and the pitch cycle T _opt from the code searching section 115 to obtain the above-described equation (2).
, A recursive pitch synthesis process is executed to generate a pitch synthesis vector Y.

Then, the gain evaluator 170 evaluates the amount of distortion between the target vector X and the pitch composite vector Y of the target vector calculator 111, and obtains a pitch gain b that minimizes the amount of distortion.
_opt is searched from the gain table 150, and information on the optimal pitch gain is provided to the drive signal generation unit 114,
The signal is output from the output terminal 119.

The drive signal generation section 114 generates a drive signal for the current subframe voice by multiplying the optimum code vector by a gain using the code information of the code search section 115 and the gain information of the gain evaluation section 170. Become like This drive signal is transmitted to the influence signal calculation unit 1.
13, the influence signal calculation unit 113 calculates an influence signal to the next subframe by encoding the current subframe for calculating a target vector of the next subframe.

Accordingly, even with the above-described coding method of the pitch synthesis filter, unlike the conventional pitch synthesis filter in which the drive signal of exactly the same shape repeats in the subframe, a recursive pitch synthesis filter is used. It is possible to cope with a sudden increase / decrease in power, and to search for a gain in the gain table only for a fixed pitch period, so that the increase in the amount of calculation required for the gain search can be negligibly small. There is also an advantage that the number of bits does not increase at all. As a result, even in speech encoding at a low bit rate with a large subframe length, a high-quality synthesized speech with a smooth power change can be obtained.

[0041]

As described above, according to the coding method of the pitch synthesis filter of the present invention, it is possible to cope with a sudden increase / decrease of the power of the drive signal, and to obtain a synthesized speech having a smooth power change. Even at a low bit rate of about 4 kbit / s, it is possible to generate high-quality synthesized speech in which the power of synthesized speech changes smoothly.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a schematic configuration of an embodiment of the present invention.

FIG. 2 is a block diagram showing a schematic configuration of a second embodiment of the present invention.

FIG. 3 is a waveform chart for explaining the effect of one embodiment of the present invention.

FIG. 4 is a diagram for explaining an encoding method of a conventional pitch synthesis filter.

[Explanation of symbols]

100: CELP encoding unit, 101: Filter coefficient calculation unit,
103: code book, 106: composite vector calculation unit,
108: pitch synthesis unit, 111: target vector calculation unit,
113: influence signal calculation unit, 114: drive signal generation unit, 1
15 code search unit, 140 synthesis unit, 150 gain table, 160 pitch synthesis unit, 162 addition unit, 1
64: Multiplication unit, 166: Delay unit, 170: Gain evaluation unit

Continuation of the front page (56) References JP-A-5-61499 (JP, A) JP-A-1-293400 (JP, A) JP-A-5-134698 (JP, A) JP-A-5-46199 (JP) JP-A-4-233000 (JP, A) JP-A-3-189700 (JP, A) JP-A-3-101800 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB Name) G10L 19/12

Claims (1)

(57) [Claims] 1. A speech encoding means for performing a pitch cycle search using an adaptive codebook and outputting a code vector of the adaptive codebook specified by the searched pitch cycle, and an output from the speech encoding means. A synthetic vector generating means for generating a synthetic vector using the code vector to be generated, and a recursive pitch synthesis filter using the pitch period and the pitch gain of the gain table determined by the adaptive code book search in the encoding means as parameters. Means for performing a recursive pitch synthesis process on a synthesized vector of the synthesized vector generating means to generate a synthesized voice, and means for selecting a pitch gain of the gain table based on a distortion amount of the synthesized voice. Characteristic coding method of pitch synthesis filter.