JP3236850B2 - Sound source vector generating apparatus and sound source vector generating method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a sound source vector generating apparatus and method for efficiently compressing / encoding / decoding audio information used in an audio encoding / decoding apparatus.

[0002]

2. Description of the Related Art CELP (Code Excited Linear Predic)
tion: "High Quality Speech at LowBit Rate", MR
Schroeder, Proc. ICASSP'85, pp. 937-940) -type speech coder performs linear prediction for each frame obtained by dividing the speech at a fixed time, and calculates a prediction residual (excitation signal) by linear prediction for each frame. ) Is encoded using an adaptive codebook storing past driving excitations and a random codebook storing a plurality of noise code vectors. The search for the adaptive codebook and the search for the random codebook are performed using the code number of the adaptive code vector and its gain (pitch gain) and the code number of the random code vector and its gain ( This is a process for determining a random code vector gain).

[0003]

(Equation 1)

Here, V is a target vector for each subframe, A is an adaptive code vector, C is a noise code vector, H is an impulse response matrix of a synthesis filter, ga is a pitch gain, and gc is a noise code vector gain. However, since the calculation cost becomes enormous when directly calculating (Equation 1), a general CELP type speech coding apparatus first performs an adaptive codebook search, and then performs a noise codebook search based on the result. Will be

Hereinafter, a search for a random codebook in the CELP type speech coding apparatus will be described with reference to FIG. In FIG. 7, X is a target vector for searching a random codebook, C
Is the noise code vector, gc is the noise code vector gain, H
(Z) is an impulse response matrix of the synthesis filter, X ′ is a vector obtained by time-reversing the target vector X and then synthesizing it with H, and then re-time-reversing the output vector thereof,
S is a vector obtained by combining C multiplied by gc with H. FIG.
In (a), X is obtained by the following (Equation 2).

[0006]

(Equation 2)

The noise codebook search is a process in which the distortion calculation unit 506 in FIG. 7A determines a code number of a noise code vector and a noise code vector gain that minimize the following (Equation 3).

[0008]

(Equation 3)

The actual CELP type speech coding apparatus employs the configuration shown in FIG. 7B in order to reduce the calculation cost. Instead of minimizing (Equation 3), the following (Equation 4) is used. The maximization is performed by the distortion calculator 611.

[0010]

(Equation 4)

On the other hand, FIG. 7C shows a noise codebook peripheral portion of the CELP type speech decoding apparatus. Actually, FIG.
And the decoding device shown in FIG. 7C are used as a pair.

[0012] Typical examples of a conventional noise codebook in a CELP type speech encoding / decoding apparatus include a plurality of random number sequences created from a random number sequence and a plurality of pulse sequence vectors stored therein. Can be However, by calculating the convolution operation result of the impulse response of the synthesis filter and the time-reversed target and the autocorrelation of the synthesis filter in advance and expanding them in the memory, the cost of calculating the coding distortion of (Equation 4) can be reduced. Since it is possible to greatly reduce the noise and to generate a noise code vector algebraically, R
Algebraic structure sound source with features such as not requiring OM ("8KBIT / S ACELP CODING OF SPEECH WITH 10 MS S
PEECH-FRAME: A CANDIDATE FOR CCITT STANDARDIZATIO
N ": R. Salami, C. Laflamme, JP. Adoul, ICASSP'94,
pp. II-97 to II-100, 1994), a CELP-type speech encoding / decoding apparatus is highly evaluated today, and a CS using the above algebraic structured sound source in a noise codebook section.
-ACELP and ACELP are respectively G.P. 729 and G.C. 723 is recommended.

[0013]

However, in the CELP-type speech coding / decoding apparatus provided with the algebraic structure excitation in the noise codebook section, the noise codebook search target is always encoded by a pulse train vector. Therefore, there is a limit in improving voice quality.

According to the present invention, it is possible to perform a search for a noise codebook at a calculation cost similar to the case where an algebraic structure excitation is used for a noise codebook, and to further generate an excitation vector capable of obtaining a high quality synthesized speech. It is intended to provide a device.

[0015]

[MEANS FOR SOLVING THE PROBLEMS]
The present invention provides an input vector supply means for supplying a pulse train vector, and a plurality of pre-created waveform patterns having a fixed waveform.
A fixed waveform storage means for storing plural as, the pulse train
According to the position and polarity sign of each pulse of the vector,
Fixed waveform arranging means for arranging a fixed waveform and outputting as a sound source vector , wherein the fixed waveform arranging means comprises
Generated by selectively using one of a number of fixed waveforms
A configuration for outputting a sound source vector is adopted.

[0016]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described with reference to FIGS.

(Embodiment 1) FIG. 1 is a block diagram showing a configuration of a sound source vector generating apparatus according to the present embodiment. In FIG. 1, reference numeral 11 denotes three fixed waveforms (V1 (length: L1), V2 (length: L1) of channels CH1, CH2, and CH3.
2), a fixed waveform storage unit for storing V3 (length: L3)),
Reference numeral 12 denotes fixed waveform start candidate position information for each channel, and the fixed waveform (V
1, V2, V3) at fixed positions P1, P2, P3, respectively. Reference numeral 13 denotes an addition unit which adds the fixed waveforms arranged by the fixed waveform arrangement unit 12 and outputs a sound source vector 14. .

The operation of the sound source vector generation device configured as described above will be described with reference to FIG. The fixed waveform storage unit 11 stores the three fixed waveforms V1, V2, and V3, and the fixed waveform arranging unit 12 determines the fixed waveform based on the fixed waveform start end candidate position information as shown in (Table 1). The fixed waveform V1 read from the storage unit 11 is arranged (shifted) at a position P1 selected from the starting candidate positions for CH1.
The fixed waveforms V2 and V3 are arranged at positions P2 and P3 selected from the starting end candidate positions for CH2 and CH3, respectively.

[0019]

[Table 1]

The adding section 13 adds the fixed waveforms arranged by the fixed waveform arranging section 12 to generate a sound source vector 14.

However, the fixed waveform starting point candidate position information held by the fixed waveform storage section 12 includes combination information of the starting point candidate positions of each selectable fixed waveform (which position is selected as P1 and which position is selected as P2). , P3), a code number corresponding to one of the positions is selected.

According to the sound source vector generation device configured as described above, it is possible to transmit voice information by transmitting a code number corresponding to the fixed waveform starting point candidate position information included in the fixed waveform arranging unit 12. At the same time, the code number is present by the product of the numbers of the starting edge candidates, so that it is possible to generate a sound source vector close to the actual voice without increasing the calculation and the required memory much.

Also, since the speech information can be transmitted by transmitting the code number, the above excitation vector generation device can be used as a noise codebook in a speech encoding / decoding device.

In this embodiment, the case where three fixed waveforms are used as shown in FIG. 1 has been described. However, the number of fixed waveforms (which corresponds to the number of channels in FIG. 1 and (Table 1)) is changed. Similar functions and effects can be obtained with other numbers.

Further, in the present embodiment, the case where fixed waveform arranging section 12 has fixed waveform starting point candidate position information shown in (Table 1) has been described. However, fixed waveform starting point candidate position information other than (Table 1) has been described. The same operation and effect can be obtained also in the case where

(Embodiment 2) FIG. 2A is a block diagram showing a configuration of a CELP type speech encoding apparatus according to the present embodiment, and FIG. 2B is a block diagram showing a configuration of a CELP type speech decoding apparatus.

In FIG. 2A, reference numeral 22 denotes a time reversing unit for time reversing the input noise codebook search target X; 23, a synthesis filter for synthesizing the output of the time reversing unit 22; A time reversing unit for reversing the time of the output of the filter and outputting a time recombining target X ′;
31 is a fixed waveform storage unit for storing a plurality of fixed waveforms, 3
2 is based on the fixed waveform start candidate position information owned by itself,
A fixed waveform arrangement unit for arranging (shifting) fixed waveforms read out from the fixed waveform storage unit 31 at selected positions;
Reference numeral 3 denotes an adding unit that adds the fixed waveforms arranged by the fixed waveform arranging unit 32 to generate the excitation vector C. Reference numeral 28 denotes a noise code vector gain gc. Reference numeral 27 denotes a synthesized excitation vector obtained by synthesizing the excitation vector C multiplied by gc. A synthesis filter that outputs S, and a distortion calculator 26 that receives X ′, C, and S and calculates distortion.

In the present embodiment, the fixed waveform storage 31
The fixed waveform arranging unit 32 and the adding unit 33 include the fixed waveform storing unit 11, the fixed waveform arranging unit 12, and the adding unit 1 shown in FIG.
3 and the starting position of the fixed waveform candidate at each channel corresponds to (Table 1). Hereinafter, the channel number, the fixed waveform number, and the symbols indicating their lengths and positions are shown in FIGS. ) Should be used.

In FIG. 2B, reference numeral 41 denotes a fixed waveform storage unit for storing a plurality of fixed waveforms, and reference numeral 42 denotes a fixed waveform storage unit 41 based on the fixed waveform starting end candidate position information of the fixed waveform storage unit 41.
A fixed waveform arrangement unit for arranging (shifting) the fixed waveforms read out from the selected positions at selected positions; 43, an addition unit for adding the fixed waveforms arranged by the fixed waveform arrangement unit 42 to generate a sound source vector C; Sign vector gain
gc and 37 are synthesis filters that synthesize the sound source vector C and output the synthesized sound source vector S.

The fixed waveform storing section 41 and the fixed waveform arranging section 42 in the speech decoding apparatus shown in FIG.
Has the same configuration as the fixed waveform storage unit 31 and the fixed waveform placement unit 32 in the speech encoding device of
The fixed waveforms stored in the equations (41) and (41) are statistically minimized by the cost function of (Equation 3) by learning using the coding distortion calculation formula of (Equation 3) as a cost function using the target for noise codebook search. It is assumed that the waveform has a fixed waveform.

The operation of the CELP-type speech encoding / decoding apparatus configured as described above will be described first with reference to FIG.

The target X for noise codebook search is time-reversed by the time reversal unit 22, then synthesized by the synthesis filter 23, time-reversed by the time reversal unit 24 again, and used for noise codebook search. The result is output to the distortion calculation unit 26 as the time reverse synthesis target X ′. Next, the fixed waveform arrangement unit 32
Is the position where the fixed waveform V1 read from the fixed waveform storage unit 31 is selected from the CH1 starting end candidate positions based on the fixed waveform starting end candidate position information of the own shown in (Table 1).
The fixed waveforms V2 and V3 are similarly arranged (shifted) at P1, and the fixed waveforms V2 and V3 are similarly selected from the starting point candidate positions for CH2 and CH3.
Place each in P3. Each of the arranged fixed waveforms is output to the adder 33 and added to generate a sound source vector C.
Through the noise code vector gain 28, the synthesis filter unit 2
7 and output to the distortion calculator 26. The synthesis filter 27 synthesizes the sound source vector C to generate a synthesized sound source vector S
Is generated and output to the distortion calculator 26.

The distortion calculator 26 receives the time inverse synthesized target X ′, the excitation vector gcC, and the synthesized excitation vector S, and calculates the encoding distortion of (Equation 4).

After calculating the distortion, the distortion calculating section 26 sends a signal to the fixed waveform arranging section 32, and the fixed waveform arranging section 32 selects the starting candidate positions corresponding to each of the three channels, and then calculates the distortion. The above processing up to the calculation of the distortion at 26 is repeated for all combinations of the starting end candidate positions that can be selected by the fixed waveform arrangement unit 32. Thereafter, the combination of the starting candidate positions where the encoding distortion is minimized is selected, the code number corresponding to the combination of the starting candidate positions one-to-one, and the noise code vector gain gc at that time, as the code of the noise codebook. Transmit to the transmission unit.

Next, the operation of the speech decoding apparatus will be described with reference to FIG.

The fixed waveform arranging section 42 selects the position of the fixed waveform in each channel from the fixed waveform starting end candidate position information shown in (Table 1) based on the information sent from the transmitting section, and The fixed waveform V1 read from the waveform storage unit 41 is arranged (shifted) at a position P1 selected from the starting end candidate positions for CH1, and similarly, fixed waveforms V2 and V3
Is a position P selected from the starting candidate positions for CH2 and CH3.
2 and P3 respectively. Each fixed waveform placed is
It is output to the adder 43 and added to become the excitation vector C, multiplied by the noise code vector gain gc selected based on the information from the transmission unit, and output to the synthesis filter unit 37. The synthesis filter 27 generates a synthesized excitation vector S by synthesizing the excitation vector C multiplied by gc, and outputs the synthesized excitation vector S.

According to the speech encoding / decoding device configured as described above, the excitation vector is generated by the excitation vector generation unit including the fixed waveform storage unit, the fixed waveform arrangement unit, and the adder. In addition to having the effect of mode 1, the synthesized sound source vector obtained by synthesizing this sound source vector with the synthesis filter has characteristics that are statistically close to the actual target, and high-quality synthesized speech can be obtained.

In this embodiment, the fixed waveform obtained by learning is stored in the fixed waveform storage units 31 and 41. In addition, the target X for noise codebook search is statistically analyzed. Similarly, when a fixed waveform created based on the analysis result is used or when a fixed waveform created based on knowledge is used, a high-quality synthesized speech can be obtained similarly.

In this embodiment, the case where the fixed waveform storage unit stores three fixed waveforms has been described.
The same operation and effect can be obtained when the number of fixed waveforms is changed to another number.

Further, in the present embodiment, the case where the fixed waveform arranging section has the fixed waveform starting point candidate position information shown in (Table 1) has been described, but it has the fixed waveform starting point candidate position information other than (Table 1). In this case, the same operation and effect can be obtained.

(Embodiment 3) FIG. 3 is a block diagram showing a configuration of a CELP type speech coding apparatus according to the present embodiment.

In FIG. 3, reference numeral 67 denotes a plurality of fixed waveforms (in this embodiment, CH1: V1, CH2: V2, C
H3: a fixed waveform storage unit for storing V3), 69 convolves the three fixed waveforms from the fixed waveform storage unit 67 with the impulse response h (length L = subframe length) of the synthesis filter, Impulse response by three types of waveforms (CH1: h
1, CH2: h2, CH3: h3, length L = subframe length), an impulse response calculation unit for each waveform, 52 is a time reversing unit for time reversing the input noise codebook search target X, 53 is a waveform-based synthesis filter that convolves the output of the time-reverse ordering unit 52 with each of the waveform-based impulse responses h1, h2, and h3 from the waveform-based impulse response calculation unit 69; And time inverse synthesis target X'1 for three waveforms
(CH1), X′2 (CH2), X′3 (CH3), a time reverse ordering unit, 59 is a fixed waveform placement unit having fixed waveform start-end candidate position information that can be generated by algebraic rules, 60
Are the start end candidate positions P1, P selected by the fixed waveform arrangement unit 59.
Only at P2 and P3, a pulse with an amplitude of 1 (with polarity) is set up, and the impulse for each channel (CH1: D1, C1
H2: D2, CH3: D3), an impulse generator 61, a noise code vector gain gc, a waveform impulse response h1, h2, h3 from a waveform impulse response calculator 69, and an autocorrelation h1, h1 And h2,
a correlation matrix calculator that calculates a cross-correlation between h1 and h3, h2 and h3, and expands the obtained correlation value in a correlation matrix memory RR;
58 is a time inverse synthesis target (X'1, X ') for three waveforms.
2, X′3), a correlation matrix memory RR, and a distortion calculation unit that calculates coding distortion of (Equation 4) using three channel-specific impulses (D1, D2, D3).

The operation of the CELP-type speech coding apparatus configured as described above will be described with reference to FIG.

First, the waveform-specific impulse response calculation unit 69 convolves the three fixed waveforms V1, V2, and V3 stored in the fixed waveform storage unit 67 with the impulse response h to obtain three types of waveform-specific impulse responses. The responses h1, h2, and h3 are calculated and output to the waveform-specific synthesis filter 53 and the correlation matrix calculator 62. Next, the waveform-specific synthesizing filter 53 convolves the noise codebook search target X time-reversed by the time reversing unit 52 with the three types of input waveform-specific impulse responses h1, h2, and h3, respectively. The time reverse ordering unit 54 time-reorders the three types of output vectors from the waveform-specific synthesis filters 53 again to generate three waveform-specific time reverse synthesis targets X′1, X′2, and X′3, respectively. Output to 58 to the calculation unit. Next, the correlation matrix calculator 62 calculates the input impulse responses h for the three types of waveforms.
The autocorrelation of each of h1, h2, h3 and h1, h2, h3
After calculating the cross-correlation between 1 and h3 and between h2 and h3, and expanding the obtained correlation value in the correlation matrix memory RR, the distortion calculation unit 58
Output to

After performing the above processing as pre-processing, the fixed waveform arranging section 59 selects starting point candidate positions of the fixed waveform for each channel one by one, and outputs the position information to the impulse generator 60. The impulse generator 60 raises a pulse having an amplitude of 1 (having a polarity) at the selected position obtained from the fixed waveform arrangement section 59 to generate the impulse D1, D for each channel.
2. D3 is generated and output to the distortion calculator 58. Then, the distortion calculation unit 58 uses the three time-wise inverse synthesis targets X′1, X′2, X′3 for each waveform, the correlation matrix memory RR, and the three impulses D1, D2, D3 for each channel to obtain 4)
Is calculated. The above processing from the selection of the starting candidate positions corresponding to each of the three channels by the fixed waveform arranging unit 59 to the calculation of the distortion by the distortion calculating unit is performed by all combinations of the starting candidate positions that the fixed waveform arranging unit can select. Is repeated. Then, a code number corresponding to the combination number of the starting end candidate position that minimizes the coding distortion of (Equation 4),
And the optimal gain at that time
Is transmitted to the transmission unit as a code of the noise codebook.

The speech decoding apparatus according to the present embodiment has the same configuration as that of the first embodiment shown in FIG.
The fixed waveform storage unit and the fixed waveform placement unit in the speech encoding device have the same configuration as the fixed waveform storage unit and the fixed waveform placement unit in the speech decoding device. The fixed waveform stored in the fixed waveform storage unit is obtained by statistically calculating the cost function of (Equation 3) by learning using the calculation formula of the coding distortion of (Equation 3) using the noise codebook search target as a cost function. It is assumed that the waveform has a fixed waveform that minimizes characteristics.

According to the speech encoding / decoding device configured as described above, the fixed waveform arranging unit has the fixed waveform starting end candidate position information that can be generated by the algebraic rule, and further includes the impulse response calculating unit for each waveform, By providing a separate synthesis filter and a correlation matrix calculator, the correlation matrix of the impulse response for each waveform and the time inverse synthesis target for each waveform can be calculated in the preprocessing stage. A search for a random codebook at the same computational cost as using an algebraic structure excitation consisting only of noise for a random codebook can be performed. Therefore, it is possible to obtain high quality synthesized speech.

In this embodiment, the case where the fixed waveform obtained by learning is stored in the fixed waveform storage unit has been described. In addition, the target X for noise codebook search is statistically analyzed and analyzed. Similarly, when using a fixed waveform created based on the result or using a fixed waveform created based on knowledge, a high-quality synthesized speech can be obtained.

Further, in the present embodiment, the case where the fixed waveform storage section stores three fixed waveforms has been described.
The same operation and effect can be obtained when the number of fixed waveforms is changed to another number.

Further, in the present embodiment, the case where the fixed waveform placement unit has the fixed waveform starting end candidate position information shown in (Table 1) has been described, but if it can be generated algebraically, (Table 1) The same operation and effect can be obtained also in the case of having the fixed waveform start candidate position information other than the above.

(Embodiment 4) FIG. 4 is a block diagram showing a configuration of a CELP type speech coding apparatus according to the present embodiment. The speech coding apparatus according to the present embodiment has two types of noise codebooks. One of the noise codebooks has the configuration of the excitation vector generation apparatus shown in FIG. Is a configuration that stores multiple random sequences,
Switching of the random codebook is performed in a closed loop.

In FIG. 4, reference numeral 81 denotes a noise codebook A, which includes a fixed waveform storage unit 82, a fixed waveform arrangement unit 83, and an addition unit 84, and corresponds to the excitation vector generation apparatus of FIG. Reference numeral 85 denotes a random codebook B, which includes a random number sequence storage unit 86 that stores a plurality of random vectors generated from a random number sequence. X is a target for searching a random codebook, 88 is a switch for switching between a random codebook A81 and a random codebook B85, 89 is a random code vector gain,
Reference numeral 90 denotes a synthesis filter that synthesizes a noise code vector output from the noise codebook connected by the switch 88, and reference numeral 91 denotes a distortion calculation unit that calculates coding distortion of (Equation 2).

The operation of the CELP-type speech coding apparatus configured as described above will be described with reference to FIG.

First, the switch 88 is connected to the noise codebook A81 side, and the fixed waveform arranging section 83 reads out from the fixed waveform storing section 82 based on the fixed waveform starting end candidate position information which the own has shown in (Table 1). The fixed waveforms are arranged (shifted) at positions selected from the starting end candidate positions. Each of the arranged fixed waveforms is output to the adder 84 and added to form a noise code vector, and is input to the synthesis filter 90 via the switch 88 and the noise code vector gain 89. The synthesis filter 90 synthesizes the input noise code vectors and outputs the result to the distortion calculation unit 91.

The distortion calculation unit 91 performs the processing of minimizing the encoding distortion of (Equation 2) using the noise codebook search target 87 and the synthesized vector obtained from the synthesis filter 90.

After calculating the distortion, the distortion calculating section 91 sends a signal to the fixed waveform arranging section 83, and from when the fixed waveform arranging section 83 selects the starting end candidate position to when the distortion calculating section 91 calculates the distortion. The above process is repeatedly performed for all combinations of the starting end candidate positions that can be selected by the fixed waveform arrangement unit 83. Thereafter, a combination of the starting candidate positions at which the encoding distortion is minimized is selected, the code number of the noise code vector corresponding to the combination of the starting candidate positions one-to-one, the noise code vector gain gc at that time, and the encoding distortion. The minimum value is stored.

Next, the switch 88 is connected to the random codebook B85 side, and the random sequence read from the random sequence storage unit 86 becomes a random code vector.
After passing through the noise code vector gain 89, the synthesis filter 90
Is input to The synthesis filter 90 synthesizes the input noise code vectors and outputs the result to the distortion calculation unit 91.

The distortion calculation unit 91 calculates the coding distortion of (Equation 2) using the noise codebook search target 87 and the synthesized vector obtained from the synthesis filter 90.

After calculating the distortion, the distortion calculating section 91 sends a signal to the random sequence storage section 86, and after selecting the random code vector by the random sequence storage section 86, the distortion calculating section 9
The above processing until the distortion is calculated in step 1 is repeated for all the random code vectors that can be selected by the random sequence storage unit 86. After that, the random code vector for which the coding distortion is minimized is selected, and the code number of the random code vector, the random code vector gain gc at that time, and the minimum coding distortion value are stored.

Next, the distortion calculating section 91 calculates the minimum coding distortion obtained when the switch 88 is connected to the noise codebook A81 and the minimum coding distortion obtained when the switch 88 is connected to the noise codebook B85. The values are compared with each other, the connection information of the switch when the smaller encoding distortion is obtained, and the code number and the noise code vector gain at that time are determined as a speech code and transmitted to the transmission unit.

The speech decoding apparatus according to the present embodiment has a configuration in which noise codebook A, noise codebook B, switch, noise code vector gain, and synthesis filter are arranged in the same configuration as in FIG. A noise codebook, a noise code vector, and a noise code vector gain to be used are determined based on a speech code input from a transmission unit, and a synthesized excitation vector is obtained as an output of a synthesis filter.

According to the speech coding / decoding device configured as described above, the noise code vector generated by the random codebook A and the noise code vector generated by the random codebook B are expressed by (Equation 2). ) Can be selected as a closed loop that minimizes the coding distortion, so that it is possible to generate a sound source vector closer to the actual speech, and to obtain a high-quality synthesized speech.

In this embodiment, the conventional CELP
Although the speech encoding / decoding device based on the configuration of FIG. 7 which is a speech encoding device of the type shown in FIG. 7, the present embodiment is applied to the CELP speech encoding / decoding device based on the configuration of FIG. 2 or FIG. Even when the embodiment is applied, the same operation and effect can be obtained.

In this embodiment, the random codebook A8
1 has the structure shown in FIG.
Has the other structure (for example, when four fixed waveforms are provided), the same operation and effect can be obtained.

In this embodiment, the random codebook A8
One fixed waveform arranging unit 83 has the fixed waveform start candidate position information shown in (Table 1), but the same operation and effect can be obtained also when the other fixed waveform start candidate position information is included.

In this embodiment, the random codebook B8
5 has been described with the random number sequence storage unit 86 that stores a plurality of random number sequences directly in the memory, but the case where the noise codebook B85 has another excitation configuration (for example, the configuration based on the algebraic structure excitation generation information The same operation and effect can be obtained for

In this embodiment, a CELP type speech coding / decoding apparatus having two types of noise codebooks has been described. However, a CELP type speech coding / decoding apparatus having three or more types of noise codebooks has been described. The same effect is obtained when
The effect can be obtained.

(Embodiment 5) FIG. 5 is a block diagram showing a configuration of a CELP type speech coding apparatus according to the present embodiment. The speech coding apparatus according to the present embodiment has two types of noise codebooks. One of the noise codebooks has the configuration of the excitation vector generation apparatus shown in FIG. Is configured by a pulse train storage unit storing a plurality of pulse trains, and adaptively switches and uses the noise codebook by using the quantization pitch gain already obtained before searching for the noise codebook.

In FIG. 5, reference numeral 101 denotes a random codebook A,
It comprises a fixed waveform storage unit 102, a fixed waveform arrangement unit 103, and an addition unit 104, and corresponds to the sound source vector generation device of FIG. Reference numeral 105 denotes a noise codebook B, which is configured by a pulse train storage unit 106 that stores a plurality of pulse trains.
107 is a noise codebook search target, 108 is a switch for switching between the random codebook A101 and the random codebook B105, 109 is a noise code vector gain, and 110 is a noise code output from one of the noise codebooks connected by the switch 108. A synthesis filter for synthesizing a code vector, 1
Reference numeral 11 denotes a distortion calculation unit that calculates coding distortion of (Equation 2), 1
12 is an adaptive codebook, 113 is a pitch gain already obtained at the time of searching for a noise codebook, and 114 is a pitch gain quantization unit.

The operation of the CELP-type speech coding apparatus configured as described above will be described with reference to FIG.

In the conventional CELP-type speech coding apparatus, a search for an adaptive codebook is first performed, and a search for a noise codebook is performed based on the search result. This adaptive codebook search is performed based on a plurality of adaptive code vectors stored in the adaptive codebook (a vector obtained by multiplying the adaptive code vector and the noise code vector by their respective gains and then adding) to an optimal adaptive codebook. This is a process of selecting a vector, and as a result, a code number and a pitch gain of an adaptive code vector are generated.

In the CELP speech coding apparatus according to the present embodiment, the pitch gain is
, And after generating a quantized pitch gain, a noise codebook search is performed. Pitch gain quantization unit 11
The quantized pitch gain obtained in 4 is sent to the switch 108 for switching the random codebook. The switch 108
When the value of the quantization pitch gain is small, it is determined that the input speech has strong voicelessness, and the noise codebook A101 is connected. When the value of the quantization pitch gain is large, the input speech is determined to be strong voiced. Then, the random codebook B105 is connected.

When the switch 108 is connected to the random codebook A101 side, the fixed waveform arranging section 103 performs the processing based on the fixed waveform start candidate position information shown in Table 1 below.
The fixed waveforms read out from the fixed waveform storage unit 102 are respectively arranged (shifted) at positions selected from the starting end candidate positions. Each of the arranged fixed waveforms is output to the adder 104 and added to form a random code vector.
8. The signal is input to the synthesis filter 110 via the noise code vector gain 109. The synthesis filter 110 synthesizes the input noise code vectors and outputs the result to the distortion calculation unit 111.

The distortion calculation section 111 calculates the coding distortion of (Equation 2) using the noise codebook search target 107 and the synthesized vector obtained from the synthesis filter 110.

After calculating the distortion, the distortion calculator 111 calculates
A signal is sent to the fixed waveform arranging unit 103 and the fixed waveform arranging unit 1
The processing from the selection of the start end candidate position to the calculation of the distortion by the distortion calculation section 111 is performed by the fixed waveform arrangement section 10
3 is repeatedly performed for all combinations of the start end candidate positions that can be selected. Thereafter, a combination of the starting candidate positions at which the encoding distortion is minimized is selected, the code number of the noise code vector corresponding to the combination of the starting candidate positions one-to-one, the noise code vector gain gc at that time, and the quantization pitch The gain is transmitted to the transmission unit as a speech code. In the present embodiment, the characteristics of unvoiced sound are reflected in advance in the fixed waveform pattern stored in the fixed waveform storage unit 102 before speech coding.

On the other hand, the switch 108 is connected to the random codebook B10
When connected to the fifth side, the pulse train read from the pulse train storage unit 106 becomes the noise code vector, and is input to the synthesis filter 110 via the switch 108 and the noise code vector gain 109. The synthesis filter 110
The input noise code vector is synthesized, and the distortion calculation unit 11
Output to 1.

The distortion calculation section 111 calculates the coding distortion of (Equation 2) using the noise codebook search target 107 and the synthesized vector obtained from the synthesis filter 110.

After calculating the distortion, the distortion calculator 111 calculates
A signal is sent to the pulse train storage unit 106, and the pulse train storage unit 1
06 selects the noise code vector and then calculates the distortion
The above process until the distortion is calculated in step 1 is repeated for all the random code vectors that can be selected by the pulse train storage unit 106. Thereafter, a noise code vector for which encoding distortion is minimized is selected, and the code number of the noise code vector, the noise code vector gain gc at that time, and the quantization pitch gain are transmitted to the transmission unit as a speech code.

The speech decoding apparatus according to the present embodiment has a configuration in which the random codebook A, the random codebook B, the switch, the random code vector gain, and the synthesis filter are arranged in the same configuration as in FIG. First, receive the transmitted quantized pitch gain, and by its magnitude,
On the encoder side, the switch 108 is connected to the noise codebook A101.
It is determined whether the connection has been made to the noise codebook B105 side. Next, a synthesized excitation vector is obtained as an output of the synthesis filter based on the code number and the sign of the noise code vector gain.

According to the excitation encoding / decoding apparatus configured as described above, the characteristics of the input speech (in the present embodiment, the magnitude of the quantization pitch gain is used as a material for determining voicedness / unvoicedness). ), The two types of noise codebooks can be adaptively switched, and when the voicedness of the input speech is strong, a pulse train is selected as a noise code vector,
When unvoicedness is strong, it becomes possible to select a noise code vector reflecting the nature of unvoiced sound, and it is possible to generate a sound source vector closer to real voicedness,
The quality of the synthesized sound can be improved. In the present embodiment, since the switching is performed in an open loop as described above, the operation and effect can be improved by increasing the amount of information to be transmitted.

In this embodiment, the conventional CELP
Although the speech encoding / decoding device based on the configuration of FIG. 7 which is a speech encoding device of the type shown in FIG. 7, the present embodiment is applied to the CELP speech encoding / decoding device based on the configuration of FIG. 2 or FIG. Even when the embodiment is applied, the same effect can be obtained.

In this embodiment, the switch 108
Was used as a parameter for switching the pitch gain of the adaptive code vector by the pitch gain quantizer 114, but a pitch period calculator was provided instead, and calculated from the adaptive code vector. A pitch cycle may be used.

In this embodiment, the random codebook A1
02 has the structure shown in FIG.
The same operation and effect can be obtained when 02 has another structure (for example, when it has four fixed waveforms).

In this embodiment, the random codebook A1
Although the fixed waveform arranging unit 103 of No. 02 has the fixed waveform start candidate position information shown in (Table 1), the same operation and effect can be obtained also in the case of having other fixed waveform start candidate position information.

In the present embodiment, the random codebook B1
A case has been described where 05 is configured by the pulse train storage unit 106 that stores the pulse train directly in the memory. However, the case where the noise codebook B105 has another excitation configuration (for example, when configured with algebraic structure excitation generation information)
The same operation and effect can be obtained for.

Although the present embodiment has been described with respect to a CELP type speech encoding / decoding apparatus having two types of noise codebooks, a CELP type speech encoding / decoding apparatus having three or more types of noise codebooks is described. The same effect is obtained when
The effect can be obtained.

(Embodiment 6) FIG. 6 is a block diagram showing a configuration of a CELP type speech coding apparatus according to the present embodiment. The speech coding apparatus according to the present embodiment has two types of noise codebooks. One of the noise codebooks has the configuration of the excitation vector generation apparatus shown in FIG. The other noise codebook has the same configuration as that of the excitation vector generating apparatus shown in FIG. 1, but the fixed waveform storage unit has two fixed waveforms. Switching between the two types of noise codebooks is performed in a closed loop.

In FIG. 6, reference numeral 121 denotes a random codebook A.
A fixed waveform storage unit A122 storing three fixed waveforms, a fixed waveform arranging unit A123, and an adding unit 124,
This corresponds to the configuration of the sound source vector generation device of FIG. 1 in which three fixed waveforms are stored in the fixed waveform storage unit. Reference numeral 131 denotes a noise codebook B, which includes a fixed waveform storage unit B132 storing two fixed waveforms, a fixed waveform placement unit B133 having fixed waveform starting end candidate position information shown in (Table 2), and a fixed waveform placement unit B133. It is composed of an adder 134 that adds the two fixed waveforms arranged to generate a noise code vector, and stores the two fixed waveforms in the fixed waveform storage unit in the configuration of the excitation vector generation device of FIG. Corresponding.

[0089]

[Table 2]

Reference numeral 144 denotes a target for searching a random codebook,
41 is a switch for switching between the random codebook A121 and the random codebook B131; 142 is a random code vector gain;
Reference numeral 43 denotes a synthesis filter that synthesizes a noise code vector output from any of the noise codebooks connected by the switch 141, and reference numeral 145 denotes a distortion calculation unit that calculates coding distortion of (Equation 2).

The operation of the CELP-type speech coding apparatus configured as described above will be described with reference to FIG.

First, the switch 145 is connected to the random codebook A12.
1 and the fixed waveform storage unit A122 is
The three fixed waveforms read out from the fixed waveform storage unit A122 are respectively arranged (shifted) at positions selected from the starting end candidate positions, based on the fixed waveform starting end candidate position information of the self indicated by (1). The arranged three fixed waveforms are output to the adder 124 and added to form a noise code vector, and are input to the synthesis filter 143 via the switch 141 and the noise code vector gain 142. The synthesis filter 143 synthesizes the input noise code vectors and outputs the result to the distortion calculation unit 145.

The distortion calculating section 145 calculates the coding distortion of (Equation 2) using the target 144 for noise codebook search and the synthesized vector obtained from the synthesis filter 143.

After calculating the distortion, the distortion calculator 145 calculates
A signal is sent to the fixed waveform arranging unit A123, and after the fixed waveform arranging unit A123 selects the starting end candidate position, the distortion calculating unit 14
The above processing up to the calculation of the distortion in step 5 is repeated for all combinations of the starting end candidate positions that can be selected by the fixed waveform arrangement unit A123. Thereafter, a combination of the starting candidate positions at which the encoding distortion is minimized is selected, the code number of the noise code vector corresponding to the combination of the starting candidate positions one-to-one, the noise code vector gain gc at that time, and the encoding distortion. The minimum value is stored. In the present embodiment, before speech coding, the fixed waveform pattern stored in fixed waveform storage unit A122 is obtained by learning such that distortion is minimized under the condition that there are three fixed waveforms. Used.

Next, the switch 145 is connected to the random codebook B131.
Side, and the fixed waveform storage unit B132 selects two fixed waveforms read from the fixed waveform storage unit B132 from the start end candidate positions based on the fixed waveform start end candidate position information included in the fixed waveform storage unit B132 shown in (Table 2). It is arranged (shifted) at each position. The arranged two fixed waveforms are output to the adder 134 and added to form a noise code vector, and are input to the synthesis filter 143 via the switch 141 and the noise code vector gain 142. The synthesis filter 143 synthesizes the input noise code vectors and outputs the result to the distortion calculation unit 145.

The distortion calculation section 145 calculates the coding distortion of (Equation 2) using the synthesized codebook search target 144 and the synthesized vector obtained from the synthesis filter 143.

After calculating the distortion, the distortion calculator 145 calculates
A signal is sent to the fixed waveform arranging unit B133.
The above processing up to the calculation of the distortion in step 5 is repeated for all combinations of the starting end candidate positions that can be selected by the fixed waveform arrangement unit B133. Thereafter, a combination of the starting candidate positions at which the encoding distortion is minimized is selected, the code number of the noise code vector corresponding to the combination of the starting candidate positions one-to-one, the noise code vector gain gc at that time, and the encoding distortion. The minimum value is stored. In the present embodiment, before speech encoding, the fixed waveform pattern stored in fixed waveform storage section B132 is obtained by learning such that the distortion is minimized under the condition that there are two fixed waveforms. Used.

Next, the distortion calculator 145 includes the switch 141
Are compared with the minimum coding distortion obtained when the switch 141 is connected to the random codebook B131, and the smaller coding distortion is obtained. The switch connection information at that time, the code number and the noise code vector gain at that time are determined as a speech code, and transmitted to the transmission unit.

The speech decoding apparatus according to the present embodiment has a structure in which the random codebook A, the random codebook B, the switch, the random code vector gain, and the synthesis filter are arranged in the same configuration as in FIG. A noise codebook, a noise code vector, and a noise code vector gain to be used are determined based on a speech code input from a transmission unit, and a synthesized excitation vector is obtained as an output of a synthesis filter.

According to the speech coding / decoding device configured as described above, the noise code vector generated by the random codebook A and the random code vector generated by the random codebook B are expressed by (Equation 2) ) Can be selected as a closed loop that minimizes the coding distortion, so that it is possible to generate a sound source vector closer to the actual speech, and to obtain a high-quality synthesized speech.

In the present embodiment, the conventional CELP
Although the speech encoding / decoding device based on the configuration of FIG. 7 which is a speech encoding device of the type shown in FIG. 7, the present embodiment is applied to the CELP speech encoding / decoding device based on the configuration of FIG. 2 or FIG. Even when the embodiment is applied, the same effect can be obtained.

In this embodiment, the random codebook A1
A case has been described in which the fixed waveform storage unit A122 stores three fixed waveforms.
Have the same number of fixed waveforms (for example, four fixed waveforms), the same operation and effect can be obtained. The same applies to the random codebook B131.

In the present embodiment, the random codebook A1
The case where the fixed waveform arrangement unit A123 of 21 has the fixed waveform starting end candidate position information shown in (Table 1) has been described.
The same operation and effect can be obtained when other fixed waveform start end candidate position information is provided. Noise codebook B131
The same applies to.

Although the present embodiment has been described with respect to a CELP-type speech encoding / decoding device having two types of noise codebooks, a CELP-type speech encoding / decoding device having three or more types of noise codebooks has been described. The same effect is obtained when
The effect can be obtained.

Further, according to the sound source vector generating apparatus of the above embodiment, it is possible to transmit voice information by transmitting a code number. / Decoding device, and can generate a sound source vector close to real speech. Also, the voice of the above embodiment
According to the encoding / decoding device, the excitation vector generation device
Low computational cost by using
Codebook search on the
Vector has characteristics that are statistically close to the actual target
Therefore, a sound source vector closer to the actual voice is generated
This makes it possible to obtain high quality synthesized speech.

[0106]

As is apparent from the above description, the present invention
According to this method, it is possible to generate a sound source vector that is closer to real speech, and to obtain a high-quality synthesized speech.

[Brief description of the drawings]

FIG. 1 is a configuration block diagram of a sound source vector generation device according to an embodiment of the present invention;

FIG. 2 is a configuration block diagram of a speech encoding / decoding device according to an embodiment of the present invention;

FIG. 3 is a configuration block diagram of a speech encoding / decoding device according to an embodiment of the present invention;

FIG. 4 is a configuration block diagram of a speech encoding / decoding device according to an embodiment of the present invention;

FIG. 5 is a block diagram illustrating a configuration of a speech encoding / decoding device according to an embodiment of the present invention;

FIG. 6 is a configuration block diagram of a speech encoding / decoding device according to an embodiment of the present invention;

FIG. 7 is a block diagram showing the configuration of a conventional CELP-type speech encoding / decoding device.

[Explanation of symbols]

11, 31, 41, 67, 82, 102 Fixed waveform storage unit 12, 32, 42, 59, 83, 103 Fixed waveform placement unit 13, 33, 43, 84, 104, 124, 134 Addition unit 14 Sound source vector 22, 52 Time reverse order 23, 27, 37, 90, 110, 143 Synthetic filter 24 Time reverse order 26, 58, 91, 111, 145 Distortion calculator 28, 38, 61, 89, 109, 142 Noise code vector gain 53 Waveform Separate synthesis filter 60 Pulse train generator 62 Correlation matrix calculator 69 Impulse response calculator for each waveform 81, 101, 121 Noise codebook A 85, 105, 131 Noise codebook B 86 Random sequence storage unit 88, 108, 141 Switch 106 Pulse train Storage section 112 Adaptive codebook 113 Pitch gain 114 Pitch gain quantizer 1 5 adaptive code vector 122 fixed waveform storage section A 123 fixed waveform arranging section A 132 fixed waveform storage section B 133 fixed waveform arranging section B

Continuation of the front page (56) References JP-A-2-280200 (JP, A) JP-A-2-282800 (JP, A) JP-A-6-130994 (JP, A) JP-A-6-195098 (JP) , A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 19/00-19/14 H03M 7/30 H04B 14/04

Claims (12)

(57) [Claims] 1. An input vector supply means for supplying a pulse train vector, and a plurality of waveform patterns created in advance as fixed waveforms.
Fixed waveform storage means for storing the number and the position and polarity code of each pulse of the pulse train vector.
Therefore, the fixed waveform is arranged and output as a sound source vector.
And a fixed waveform arranging means for force, said fixed waveform arranging means are all of the plurality of fixed waveforms
Output the sound source vector generated by selectively using
Excitation vector generating apparatus characterized by that. 2. The method according to claim 1 , further comprising voicedness determination means.
A sound with a fixed waveform arranged according to the judgment result of the judgment means
Either the source vector or the pulse train vector
2. The method according to claim 1, wherein the output is performed as a vector.
Sound source vector generation device. 3. A fixed waveform and a pulse column vector, wherein the placement of fixed waveforms by convolving the claim 1 or claim 2 excitation vector generator according. 4. An input vector supply means for supplying a pulse train vector, and a plurality of waveform patterns created in advance as fixed waveforms.
Fixed waveform storage means for storing the number and the position and polarity code of each pulse of the pulse train vector.
Therefore, the fixed waveform is shifted, and
Fixed waveform shifting means for outputting , wherein the fixed waveform shifting means includes any one of the plurality of fixed waveforms.
Output sound source vector generated by selectively using
A sound source vector generation device. 5. The method according to claim 1 , further comprising voicedness determination means.
The fixed waveform has been shifted according to the judgment result of the judgment means.
Either the sound source vector or the pulse train vector
5. The method according to claim 4, wherein the output is performed as a vector.
Sound source vector generator. 6. The sound source vector generation apparatus according to claim 4 , wherein the fixed waveform is shifted by convolving the fixed waveform with a pulse train vector. 7. A step of supplying a pulse train vector, and using a plurality of waveform patterns created in advance as fixed waveforms.
The method comprising several feed, to the position and polarity sign of each pulse of the pulse train vector
Therefore, the fixed waveform is arranged and output as a sound source vector.
Fixing the fixed waveforms, wherein the fixed waveforms are arranged in any one of the plurality of fixed waveforms.
Output the sound source vector generated by selectively using
Excitation vector generating method, characterized in that that. 8. The method according to claim 8 , further comprising the step of determining voicedness.
A sound with a fixed waveform arranged according to the judgment result of the judgment stage
Either the source vector or the pulse train vector
8. The method according to claim 7, wherein the output is performed as a vector.
Sound source vector generation method. 9. The sound source vector generation method according to claim 7 , wherein the fixed waveform is arranged by convolving the fixed waveform with a pulse train vector. 10. A step of supplying a pulse train vector, and a step of combining a plurality of waveform patterns created in advance as fixed waveforms.
The method comprising several feed, to the position and polarity sign of each pulse of the pulse train vector
Therefore, the fixed waveform is shifted, and
Outputting a fixed waveform shift step, wherein the fixed waveform shift step includes any of the plurality of fixed waveforms.
Output sound source vector generated by selectively using
A sound source vector generating method. 11. The method according to claim 11 , further comprising the step of determining voicedness.
The fixed waveform is shifted according to the judgment result of the gender judgment stage.
Sound source vector or pulse train vector
11. The output as a source vector.
The described sound source vector generation method. 12. The sound source vector generation method according to claim 10 , wherein the fixed waveform is shifted by convolving the fixed waveform with a pulse train vector.