JP3297749B2 - Encoding method - Google Patents

Abstract

PURPOSE:To enable encoding of high quality by finding on-frequency-axis data in block units of an audio signal, and switching and quantizing plural code books corresponding to parameters showing features of the respective blocks. CONSTITUTION:A frequency axis conversion processing part 12 provided with a block division part 12a which divides the input speech signal, etc., on the time base into blocks of a specific number of samples, an orthogonal conversion part 12b, a data processing part 12c for finding amplitude information representing the features of a spectrum envelope, etc., converts the speech signal, etc., into the spectrum amplitude data on the frequency axis. A vector quantization part 15 puts a specific number of input data into M-dimensional vectors and thus performs vector quantization. This vector quantization part 15 is equipped with plural kind of code books and switches the code books according to the parameters showing the features of the blocks of the input signal to quantize the signal. Consequently, the vector quantization is performed efficiently corresponding to the characteristics, etc., of the input data.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an encoding method for dividing an input audio signal such as a voice signal and an audio signal into frames, converting the signal into data on a frequency axis, and performing encoding.

[0002]

2. Description of the Related Art There are known various encoding methods for compressing an audio signal (including a voice signal and an acoustic signal) by utilizing a statistical property in a time domain and a frequency domain and a characteristic of human perception. ing. As this encoding method,
Coding in the time domain, coding in the frequency domain,
Analysis synthesis coding.

[0003] As an example of high-efficiency encoding of a speech signal or the like, M
BE (Multiband Excitation) coding, SBE (Singleband Excitation) coding, Harmonic coding, SBC
(Sub-band Coding: band division coding), LPC (Linear
Predictive Coding: Linear predictive coding) or DC
T (Discrete Cosine Transform), MDCT (Modified DC
Conventionally, when quantizing various information data such as spectrum amplitude and its parameters (LSP parameter, α parameter, k parameter, etc.) in T), FFT (Fast Fourier Transform), scalar quantization is conventionally performed. There are many.

[0004]

By the way, if the bit rate is reduced to, for example, about 3 to 4 kbps and the quantization efficiency is further improved, the quantization noise (distortion) increases in the scalar quantization. It was difficult to put it to practical use. Therefore, the time axis data, the frequency axis data, the filter coefficient data, and the like obtained at the time of encoding are not individually quantized, but a plurality of data are grouped into a set (vector) and expressed by one code. Vector quantization for quantization is attracting attention.

[0005] Here, the spectral envelope of the above MBE, SBE, LPC, etc., or its parameters (L
At the time of vector quantization of the SP parameter, α parameter, k parameter, and the like, a fixed codebook is used. However, if the number of usable bits decreases (lower bit rate), sufficient performance cannot be obtained with a fixed codebook. For this reason, it is preferable to perform vector quantization on clustering (classification) of input data to be vector-quantized so that the existing area in the vector space is reduced.

[0006] Further, even when the transmission bit rate has a margin, it is considered to use a structured codebook in order to reduce the codebook size and the amount of calculation for search. Output index length is n
Instead of using one codebook of +1 bits,
For example, it is preferable to divide the codebook into two n-bit codebooks.

The present invention has been made in view of such circumstances, and it is possible to efficiently perform vector quantization in accordance with the characteristics of input data and the like, and to calculate the size of the codebook of the vector quantizer and the calculation at the time of search. An object of the present invention is to provide an encoding method that can reduce the amount and perform high-quality encoding.

[0008]

According to the high efficiency coding method of the present invention, an input audio signal (speech signal, audio signal, etc.) is divided into blocks and converted into data obtained by converting the data into a frequency axis. A step of obtaining data on the frequency axis as an M-dimensional vector based on the plurality of codebooks corresponding to the state of the audio signal in order to perform vector quantization on the data on the frequency axis of the M-dimensional vector And a step of performing quantization by switching the plurality of codebooks according to a parameter representing a feature of each block of the audio input signal using a quantizer.

According to another aspect of the present invention, there is provided a high-efficiency encoding method which divides an input audio signal into blocks and converts the data into a frequency axis. Obtaining the M-dimensional vector, dividing the data on the frequency axis of the M-dimensional vector into a plurality of groups, and obtaining a representative value for each group to reduce the dimension to S-dimensional (S <M); Performing a first vector quantization on the data; dequantizing the first vector quantization output data to obtain a corresponding S-dimensional code vector;
Extending the S-dimensional code vector to the original M-dimensional vector; and performing second processing on the data representing the relationship between the expanded M-dimensional vector and the original data on the frequency axis of the M-dimensional vector. Using a second vector quantizer having a plurality of codebooks according to the state of the audio signal to perform the vector quantization of the plurality of codes in accordance with a parameter representing a feature of each block of the audio input signal. The above-mentioned problem is solved by having a step of performing quantization by switching a book.

In these inventions, when an audio signal is used as the audio signal, a plurality of codebooks according to whether the audio signal is voiced or unvoiced are used as the codebook, and the input for each block is used as the characteristic parameter. A parameter indicating whether the audio signal is voiced or unvoiced can be used. Also, as feature parameters,
The pitch value, the strength of the pitch component, the voiced / unvoiced sound content ratio, the slope of the signal spectrum and its level, etc. can be used, and it is basically preferable to switch the codebook according to the voiced or unvoiced sound. Such a characteristic parameter may be transmitted separately, or may be replaced with an originally transmitted parameter defined in advance by an encoding method. Further, as the data on the frequency axis of the M-dimensional vector, data obtained by nonlinearly compressing the data converted to the frequency axis in the block unit can be used. Further, before the vector quantization, a difference between blocks of data to be vector-quantized is calculated,
Vector quantization may be performed on the inter-block difference data.

[0011]

By performing vector quantization by switching the plurality of codebooks according to parameters representing characteristics of each block of the input audio signal, quantization can be performed efficiently, and the codebook of the vector quantizer can be used. The size and the amount of calculation at the time of search can be reduced, and high-quality encoding can be performed.

An embodiment of the encoding method according to the present invention will be described below with reference to the drawings. FIG. 1 shows a schematic configuration of an encoding device (encoder) for explaining a high-efficiency encoding method which is an embodiment of the encoding method according to the present invention.

In FIG. 1, a voice signal or an acoustic signal is supplied to an input terminal 11, and this input signal is converted into spectrum amplitude data on a frequency axis by a frequency axis conversion processing unit 12. Inside the frequency axis conversion processing unit 12, for example, a blocking unit 1 for blocking every predetermined number of samples (N samples) of the input signal on the time axis
2a, orthogonal transform unit 12 such as FFT (fast Fourier transform)
b, a data processing unit 12c for obtaining amplitude information representing characteristics of the spectrum envelope is provided. An output from the frequency axis conversion processing unit 12 is sent to a vector quantization unit 15 via a non-linear compression unit 13 which converts the data into, for example, a dB region as needed, and a processing unit 14 which takes a difference between blocks as needed. Sent. This vector quantization unit 1
In step 5, a predetermined number (M samples) of input data is combined into M dimensions and vectors, and vector quantization processing is performed. In such an M-dimensional vector quantization process, generally, a code vector having the closest distance in an M-dimensional space to an input M-dimensional vector is searched from a codebook, and the search is performed. This is a process of extracting the index of the code vector obtained from the output terminal 16, but the vector quantization unit 15 of the embodiment shown in FIG.
Has a plurality of types of codebooks, and these codebooks can be switched in accordance with parameters representing characteristics of the input signal from the frequency axis conversion processing unit 12.

In the example of FIG. 1, an audio signal is assumed as an input signal, and a codebook 15 _V for V (voiced sound) and UV
(Unvoiced) codebook 15 _U and changeover switch 15
_{The signal is} switched by _W and sent to the vector quantizer 15 _{Q. The} changeover switch 15 is switched and controlled in accordance with a V / UV (voiced sound / unvoiced sound) discrimination signal from the frequency axis conversion processing unit 12. You. This V / UV discrimination signal (flag or the like) is transmitted to an MBE (Multiband Ex
In the case of a citation (multi-band excitation) vocoder (speech analysis / synthesis device) or the like, the parameter is transmitted from the analysis system (encoder) to the synthesis system (decoder), and does not need to be separately transmitted.

Here, the case of the MBE will be described as an example. The parameter for switching between the codebooks 15 _V and 15 _U is V / UV which is one of the information to be transmitted.
What is necessary is just to use a discrimination flag. That is, the frequency axis conversion processing unit 12 performs band division according to the pitch, and performs V (voiced sound) for each of the divided bands.
Or UV (unvoiced sound) is determined. Here, the number of V bands is N _V , the number of UV bands is N _UV, and for a predetermined threshold V _th ,

[0016]

(Equation 1) In this case, the codebook 15 _V for V (voiced sound) is switched and selected. Otherwise, the codebook 15 _U for UV (unvoiced sound) is switched and selected. The threshold value V _th may be set to, for example, about 1.

Similarly, the decoder (synthesis system) side selects and switches between two types of codebooks for V (voiced sound) and UV (unvoiced sound). Since the V / UV discrimination flag is side information (auxiliary information) that is always transmitted in the MBE vocoder, it is not necessary to separately transmit a characteristic parameter for codebook switching in this specific example, and the transmission bit rate is increased. There is no.

Codebook for V (voiced sound) 15 _V , UV
Generating codebook 15 _U (unvoiced sound) (training) is simply made possible by dividing the training data on the same basis. That is, a codebook generated from a set of amplitude data determined as V (voiced sound) is referred to as a _V codebook 15 _V, and a codebook generated from a set of amplitude data determined as UV (unvoiced sound) is referred to as UV. Code book 15 _U.

In this specific example, since the V / UV information is used for switching the codebook, the V / UV information is used.
It is necessary to make the determination flag more reliable (higher reliability). For example, in a part that can be clearly regarded as a consonant or background noise, the whole band UV should be used. As an example of the determination, it is possible to convert a small input having a large power in a high frequency range to UV.

FFT (Fast Fourier Transform) is performed on N points (256 samples) of the input signal, and valid 0 to π
In (0 to N / 2), power calculation is performed in each section of 0 to N / 4 and N / 4 to N / 2.

[0021]

(Equation 2) Using P _L and P _H of the equation (2),

[0022]

(Equation 3) When Rd <R _th and L <L _th , it is unconditionally determined to be all band UV (unvoiced sound).

By doing so, when an incorrect pitch is detected by a minute input, there is an effect that it is not used. Producing a more reliable V / UV flag in this way is convenient for switching codebooks during vector quantization.

Next, training for creating the codebooks for V (voiced sound) and UV (unvoiced sound) will be described with reference to FIG. In FIG.
A signal from a training set 31 composed of a voice signal for about several minutes for training is applied to a frequency axis conversion processing unit 32.
The pitch extraction is performed by the pitch extraction unit 32a.
The spectrum amplitude is calculated by the spectrum amplitude calculation unit 32b, and the V / UV determination for each band is performed by the V / UV determination unit 32c for each band. The output data from the frequency axis conversion processing unit 32 is sent to the training pre-processing unit 34.

In the training pre-processing unit 34, V / UV
The content check unit 34a calculates the above equation (1) or (4)
The condition of the formula is checked, and the above-mentioned spectrum amplitude data is distributed by the training data distribution unit 34b according to the obtained V / UV information.
In the case of, the training data output unit 3 for V (voiced sound)
6a, and in the case of UV (unvoiced sound), the amplitude data is sent to the training data output unit 37a for UV (voiced sound).

The spectrum amplitude data of V (voiced sound) output from the V training data output unit 36a is sent to a training processing unit 36b, for example, so-called L
Training processing is performed by the BG method, and V (voiced sound)
The codebook 36c is created. Here, the above LB
What is the G method? Linde, Y., Buzo, A. and Gray, RM, "An
Algolithm for Vector Quantizer Design ", IEEE Tran
s. Comm., COM-28, pp.84-95, Jan. 1980, is a codebook training method for an algorithm for designing a vector quantizer. On the other hand, a substantially optimal vector quantizer is designed using a training sequence. Similarly, the UV training data output unit 37
The UV (unvoiced sound) spectral amplitude data output from a is sent to the training processing unit 37b, for example, LB
Training process is performed by G method, UV (unvoiced sound)
The codebook 37c is created.

Here, as will be described later, the vector quantization unit has a hierarchical structure, the upper layer uses a V / UV shared codebook, and only the lower layer codebook is V / UV.
When switching according to UV, V / UV
You also need to create a codebook for the common part. In this case, the output data from the frequency axis conversion
It is necessary to send the / UV shared part to the codebook training data output unit 35a.

The spectrum amplitude data output from the codebook training data output unit 35a for the V / UV common part is sent to the training processing unit 35b, where the training processing is performed by, for example, the so-called LBG method, and the V / UV common use is performed. A partial codebook 35c is created. The generated code vector from the V / UV shared code book 35c is sent to the V and UV training data output units 36a and 37a.
Training for V and UV using codebook
Performs upper layer vector quantization on the data,
Generate V and UV training data for
Need to be

Hereinafter, the specific structure and operation of the hierarchically structured vector quantization unit will be described with reference to FIGS. In other words, the vector quantization unit 15 shown in FIG. 3 has, for example, a hierarchical structure in two layers above and below, and performs two-stage vector quantization on an input vector.

An input terminal 17 of the vector quantizer 15 shown in FIG. 3 receives the amplitude data on the frequency axis from the frequency axis conversion processing unit 12 shown in FIG. Via the difference processing unit 14), and supplied as the above-described M-dimensional vector as a unit of vector quantization. The M-dimensional vector is divided into a plurality of groups by being sent to the dimension reduction unit 21, and the dimension is reduced to S-dimension (S <M) by obtaining a representative value for each group. Here, FIG.
Shows a specific example of each element of the M-dimensional vector X input to the frequency domain, that is, M pieces of amplitude data x (n) on the frequency axis, where 1 ≦ n ≦ M. These M pieces of amplitude data x
(n) is collected, for example, every four samples to obtain a representative value, for example, an average value y _i , and as shown in FIG. 5, _S data of average value data y _{1 to} y S (in this case, S = M
/ 4) is obtained.

Next, vector quantization is performed on the data of the S-dimensional vector by the S-dimensional vector quantizer 22 _Q. That is, the S-dimensional vector quantizer 22 _Q
Of S dimension code vectors of the codebook 22 in _C,
A code vector that is closest to the input S-dimensional vector in the S-dimensional space is searched. Index data of the searched code vector is extracted from the output terminal 26, and the searched code vector (the output index is set to the inverse vector). The code vector obtained by quantization is sent to the dimension extension unit 23. Codebook 22 _C
The code book 35c of the V / UV shared part described with reference to FIG. 6, after the vector quantization of the S-dimensional vector Y consisting of S avg value data y ₁ ~y _S shown in FIG. 5, the quantum by inversely quantizing (or codebook of the vector quantizer 22 Vector code Y _VQ as a local decoder output obtained by extracting the code vector searched at the time of conversion
Shows each element y _VQ1 ~y _VQS of.

Next, the dimension extension unit 23 extends the S-dimensional code vector to the original M-dimensional vector. FIG. 7 shows an example of each element of the extended M-dimensional vector.
As apparent from FIG. 7, by increasing each element y _VQ1 ~y _VQS of the inverse vector quantized S-dimensional vector Y _VQ in each respective original 4 samples, 4
An M-dimensional vector consisting of S = M elements is obtained. The second vector quantization is performed on the data representing the relationship between the expanded M-dimensional vector and the original data on the frequency axis of the M-dimensional vector.

In the specific example of FIG. 3, the extended M-dimensional vector data from the dimension extension unit 23 is sent to the subtractor 24,
By subtracting the original M-dimensional vector from the data on the frequency axis, S vector data representing the relationship between the S-dimensionally expanded M-dimensional vector and the original M-dimensional vector is obtained. FIG. 8 shows M amplitude data x on the frequency axis which are each element of the M-dimensional vector X shown in FIG.
from (n), shows the M data r ₁ ~r _M obtained by subtracting each element of the extended M-dimensional vector shown in FIG. 7,
The four samples of the _M data r _{1 to} r M are set as a set (vector) and S four-dimensional vectors R _{1 to} R _M are set.
_S is obtained.

The S obtained from the subtractor 24 in this manner
Each of the vectors is subjected to vector quantization by each of the S vector quantizers 25 _{1Q to} 25 _SQ of the vector quantizer group 25. Each vector quantizer 2
5 index output from _1Q to 25 _SQ are respectively taken from the output terminal 27 _1Q ~ 27 _SQ. FIG. 9 shows a four-dimensional vector after quantizing each of the four-dimensional vectors R _{1 to} R _S shown in FIG. 8 using four-dimensional vector quantizers as the vector quantizers 25 _{1Q to} 25 _SQ . R _VQ1 ~R each element of the _VQS r _VQ1 ~r
_VQ4, r _VQ5 ~r _VQ8, shows the ... ~r _VQM.

Each of these vector quantizers 25 _{1Q to} 25 _Q
Each _SQ has a codebook for V (voiced sound) 25 _1V
~ 25 Codebook for _SV and UV (unvoiced sound) 25 _1U ~ 2
5 _SU, and these V codebooks 25 _1V-
The 25 _SV and the UV codebooks 25 _{1U to} 25 _SU are switched and selected by the changeover switches 25 _{1W to} 25 _SW which are switched and controlled according to the V / UV information from the input terminal 18. Switching control of these changeover switches 25 _1W to 25 _SW may be performed simultaneously (in conjunction) with respect to all the bands, but considering that the frequency bands responsible each vector quantizer 25 _1Q to 25 _SQ is different And V for each band
Switching control may be performed according to the / UV determination flag. The V codebooks 25 _{1V to} 25 _SV correspond to the V (voiced sound) codebook 36c in FIG. 2, and the UV codebooks 25 _{1U to} 25 _SU correspond to the V (voiced sound) codebook 37c in FIG. Of course you do.

By performing such hierarchically structured two-stage vector quantization, the amount of calculation for a codebook search can be reduced, and the amount of memory (eg, ROM capacity) for a codebook can be reduced. The output terminal 2
For example, by applying error correction coding to the more important index of the upper layer obtained from No. 6 to protect the index more importantly, it is possible to effectively apply the error correction code. Note that the hierarchical structure of the vector quantization unit 15 is not limited to two levels, and may have a multilayer structure of three or more levels.

It is not necessary that all of the units shown in FIGS. 1 to 3 be constituted by hardware.
(Digital signal processor) or the like may be used as software.

As described above, in the case of speech synthesis analysis coding, for example, the degree of voiced / unvoiced speech, the pitch, and the like are already taken out as feature amounts, and these feature amounts, particularly voiced, are considered. By switching the codebook of vector quantization according to the / unvoiced determination result, good vector quantization can be realized. In other words, the shapes of existing spectra are significantly different between voiced sound and unvoiced sound, and it is very preferable to have codebooks that are individually trained corresponding to each state in terms of improving characteristics. In the case of hierarchical structured vector quantization, the upper layer vector quantization may be a fixed codebook, and only the lower layer vector quantization codebook may be switched between voiced and unvoiced. . In addition, the bit allocation on the frequency axis may be switched to, for example, emphasizing low sounds for voiced sounds and emphasizing high sounds for unvoiced sounds. Switching control includes the presence or absence of pitch,
The ratio of voiced / unvoiced sound, spectrum level and inclination, etc. can be used. Further, three or more codebooks may be switched. For example, two or more codebooks for unvoiced sounds may be used for consonants and background noise.

Next, a specific example of a kind of MBE (Multiband Excitation) vocoder which is a kind of speech signal synthesis / analysis coding apparatus (so-called vocoder) to which the above-described high efficiency coding method can be applied will be described. This will be described with reference to the drawings. The MBE vocoder described below is compatible with DW Griffin and JS Lim, "MultibandExcitatio
n Vocoder, "IEEE Trans.Acoustics, Speech, and Signal
Processing, vol. 36, No. 8, pp. 1223-1235, Aug. 1988
And the conventional PARCOR (PA
Rtial auto-CORrelation: In vocoders and the like, voiced sections and unvoiced sections are switched for each block or frame when modeling speech, whereas MBE vocoders use the same time (the same block or the same block). The modeling is performed on the assumption that a voiced section (Voiced) section and an unvoiced section (Unvoiced) section exist in the frequency axis region (in a frame).

FIG. 10 is a block diagram showing an overall schematic configuration of an embodiment in which the present invention is applied to the MBE vocoder. In FIG. 10, an audio signal is supplied to an input terminal 101, and this input audio signal is
The signal is sent to a filter 102 such as an HPF (high-pass filter) to remove a so-called DC (direct current) offset and to remove at least a low-frequency component (200 Hz or less) for band limitation (for example, limited to 200 to 3400 Hz). . The signal obtained through the filter 102 is sent to the pitch extraction unit 103 and the windowing processing unit 104, respectively.
In the pitch extracting section 103, the input audio signal data is divided into blocks (or cut out by a rectangular window) in units of a predetermined number N (for example, N = 256), and the pitch of the audio signal in this block is extracted. .
For example, as shown in FIG. 11A, an L sample (for example, L = 1)
60) is moved in the time axis direction at the frame interval,
The overlap between the blocks is NL samples (for example, 96 samples). In addition, the windowing processing unit 10
4, a predetermined window function for one block N samples,
For example, a hamming window is applied, and the windowed block is sequentially moved in the time axis direction at intervals of L samples of one frame.

When such a windowing process is represented by a mathematical formula, _xw (k, q) = x (q) w (kL-q) (4) In the equation (4), k is a block number,
q represents a time index (sample number) of the data. The q-th data x (q) of the input signal before processing is windowed by a window function w (kL-q) of the k-th block. This shows that data x _w (k, q) can be obtained. The window function w _r (r) in the case of a rectangular window as shown in FIG. 11A in the pitch extraction unit 103 is expressed as: w _r (r) = 1 0 ≦ r <N (5) = 0 r <0, N ≦ r Further, the window function w _h (r) in the case of the Hamming window as shown in FIG. 11B in the windowing processing unit 104 is represented by w _h (r) = 0.54−0.46 cos (2πr / (N-1)) 0 ≦ r <N (6) = 0 r <0, N ≦ r. When such a window function w _r (r) or w _h (r) is used, the window function w (r) (= w (kL−
zero data interval q)) is deformed this 0 ≦ kL-q <N, kL-N < window function in the case of q ≦ kL Thus for example the rectangular window _{w r (kL-q) =} 1
As shown in FIG. 12, kL−N <q ≦ kL
It is time of. Further, the above equations (4) to (6) indicate that the length N
(= 256) indicates that the sample window advances by L (= 160) samples. Hereinafter, the above (5)
Each N point (0 ≦ r) extracted by each window function of the equation (6)
<N) are represented by x _wr (k, r) and x _wh , respectively.
(k, r).

As shown in FIG. 13, the windowing processing unit 104 obtains 17 samples from the sample sequence x _wh (k, r) of 256 samples per block to which the Hamming window of the above equation (6) has been applied.
Zero data for 92 samples is added (so-called zero padding) to make 2048 samples, and the orthogonal transform unit 105 performs orthogonal transform such as FFT (fast Fourier transform) on the time axis data sequence of 2048 samples. Processing is performed.

In the pitch extracting section 103, the above x _wr (k, r)
Is extracted based on the sample sequence (1 block N samples). As the pitch extraction method, a method using a periodicity of a time waveform, a periodic frequency structure of a spectrum, an autocorrelation function, and the like are known.
The autocorrelation method of the center clip waveform is adopted. As for the center clip level in the block at this time, one clip level may be set for each block, but the peak level of the signal of each part (each sub-block) obtained by subdividing the block is detected, and these are detected. When the difference between the peak levels of each sub-block is large, the clip level in the block is changed stepwise or continuously. The pitch period is determined based on the peak position of the autocorrelation data of the center clip waveform. At this time, a plurality of peaks are obtained from the autocorrelation data belonging to the current frame (the autocorrelation is obtained from data of one block N samples), and the maximum peak among the plurality of peaks is equal to or larger than a predetermined threshold. In this case, the maximum peak position is used as the pitch cycle, and in other cases, a pitch within a pitch range that satisfies a predetermined relationship with a pitch obtained in a frame other than the current frame, for example, the previous and next frames, for example, the center of the pitch of the previous frame. As a result, a peak within a range of ± 20% is obtained, and the pitch of the current frame is determined based on the peak position. In the pitch extracting section 103, a relatively rough pitch search is performed by an open loop, and the extracted pitch data is stored in a high-precision (fine) pitch searching section 10
6 to perform a high-precision pitch search (fine search of pitch) by a closed loop.

High-precision (fine) pitch search section 106
Contains the coarse (rough) pitch data of the integer value extracted by the pitch extracting unit 103 and the orthogonal transform unit 10.
5, for example, the data on the frequency axis that has been subjected to the FFT. In the high precision pitch search unit 106,
With the coarse pitch data value as the center, ± 0.2 steps
Shake several samples at a time to drive to the optimal fine pitch data with a decimal point (floating). At this time, as a method of fine search, a so-called analysis by synthesis method is used, and the pitch is selected so that the synthesized power spectrum is closest to the power spectrum of the original sound.

The fine search of the pitch will be described. First, in the MBE vocoder, the FF
S (j) as spectrum data on the frequency axis orthogonally transformed by T or the like is expressed as S (j) = H (j) | E (j) | 0 <j <J (7) Model is assumed. Where J is
_{ω s / 4π = f s /} 2 in response, the sampling frequency f
_s = when ω _s / 2π is, for example, 8kHz corresponding to 4kHz. In the above equation (7), when the spectrum data S (j) on the frequency axis has a waveform as shown in FIG.
H (j) indicates the spectral envelope of the original spectral data S (j) as shown in FIG.
E (j) indicates the spectrum of the excitation signal (excitation) at the same level and periodic as shown in FIG. 14C. That is, the FFT spectrum S (j) is represented by the spectrum envelope H (j) and the power spectrum of the excitation signal |
E (j) |

Power spectrum | E (j) of the above excitation signal
| Takes into account the periodicity (pitch structure) of the waveform on the frequency axis determined according to the pitch, and converts the spectrum waveform corresponding to the waveform of one band (band) for each band on the frequency axis. It is formed by arranging it repeatedly. The waveform for one band is obtained by performing a FFT by regarding a waveform obtained by adding 0 data of 1792 samples (0 padding) to a Hamming window function of 256 samples as shown in FIG. It can be formed by cutting out an impulse waveform having a certain bandwidth on the frequency axis according to the pitch.

Next, for each band divided according to the pitch, a value (a kind of amplitude) | A representative of the above H (j) (to minimize the error for each band) | A _m
| Here, for example, when the lower and upper points of the m-th band (band of the m-th harmonic) are a _m and b _m , respectively, the error ε _m of the m-th band is

[0048]

(Equation 4)

Can be expressed by | A _m | that minimizes this error ε _m is

[0050]

(Equation 5)

When | A _m | in equation (6),
Minimize the error ε _m . Such an amplitude | A _m | is obtained for each band, and an error ε _m for each band defined by the above equation (8) is obtained using the obtained amplitudes | A _m |. Next, a total value Σε _m of all the bands of the error ε _m for each band is obtained. Further, the error sum Shigumaipushiron _m of all such bands, calculated for different pitches to some small, obtaining the pitch as an error sum Shigumaipushiron _m is minimized.

That is, several types are prepared vertically, for example, in increments of 0.25 with the rough pitch obtained by the pitch extraction unit 103 as the center. Each respective pitches of different pitches to these plurality of types of fine finding the error sum Σε _m. In this case, when the pitch is determined, the bandwidth is determined. From the above equation (9), the power spectrum | S (j) | of the data on the frequency axis and the excitation signal spectrum | E
(j) | seek error epsilon _m of equation (8) using a, can be obtained sum Shigumaipushiron _m of all the bands. Obtains the error sum Shigumaipushiron _m for each pitch, is not to determine the pitch corresponding to error sum total value which is the smallest as the optimal pitch. As described above, the optimum fine (for example, in increments of 0.25) pitch is obtained by the high-precision pitch search unit 106, and the amplitude | A _m |
Is determined.

In the above description of the fine search of the pitch, in order to simplify the description, all the bands are voiced (Vo).
iced), but the MBE
In a vocoder, an unvoiced sound (Un
Since the model in which a voiced) region exists is used, it is necessary to discriminate voiced / unvoiced sounds for each band.

The data of the optimum pitch and amplitude | A _m | from the high-precision pitch search unit 106 is sent to the voiced / unvoiced sound discriminating unit 107, and the voiced / unvoiced sound is discriminated for each band. For this determination, NSR (noise-to-signal ratio) is used. That is, the NSR of the m-th band is

[0055]

(Equation 6)

When the NSR value is larger than a predetermined threshold value (for example, 0.3) (error is large), | S (j) by | A _m || E (j) | | Is not good (the excitation signal | E (j) | is inappropriate as a basis), and the band is identified by UV (Unvoice
d, unvoiced sound). In other cases, it can be determined that the approximation has been performed to some extent, and the band is
(Voiced, voiced sound).

Next, the amplitude re-evaluation unit 108 receives the on-frequency data from the orthogonal transformation unit 105 and the amplitude | A _m evaluated as a fine pitch from the high-precision pitch search unit 106.
| And the voiced / unvoiced sound discriminating unit 107
V / UV (voiced sound / unvoiced sound) discrimination data is supplied. The amplitude reevaluating unit 108 calculates the amplitude again for the band determined to be unvoiced (UV) by the voiced / unvoiced sound determining unit 107. The amplitude | A _m | _UV for this UV band is

[0058]

(Equation 7)

Is obtained by

The data from the amplitude reevaluating unit 108 is
Data number conversion (a kind of sampling rate conversion) unit 10
9 The data number conversion unit 109 is provided to make the number constant in consideration of the fact that the number of division bands on the frequency axis differs according to the pitch and the number of data (particularly the number of amplitude data) differs. is there. That is, for example, when the effective band is up to 3400 kHz, this effective band is divided into 8 to 63 bands according to the pitch, and the amplitude | obtained for each of these bands is obtained.
A _m | will change _(UV including UV band amplitude | | A _m) the number of data m _MX +1 also 8-63. Therefore, the data number conversion unit 109 converts the variable number m _MX +1 of amplitude data into a fixed number N _C (for example, 44) data.

In this embodiment, dummy data is used to interpolate values from the last data in the block to the first data in the block with respect to the amplitude data of one effective band on the frequency axis. Is added to expand the number of data to N _F , and then the band-limited K _OS times (for example, 8
Obtain an amplitude data of K _OS times the number by performing oversampling multiplied), the K _OS times the number ((m _MX +
1) × K _OS amplitude data is linearly interpolated and expanded to more N _M (for example, 2048), and this N _M data is decimated to obtain the constant number N _C (for example, 44). Convert to data.

The data (the fixed number N _C of amplitude data) from the data number conversion unit 109 is sent to the vector quantization unit 110 and is grouped into a predetermined number of data to form a vector. Will be applied. As the vector quantization unit 110, FIG.
, A vector quantization unit 15 for switching and selecting between a codebook for V (voiced sound) and a codebook for UV (unvoiced sound) is used. These V and UV codebooks are stored in the voiced / unvoiced (V / UV) discriminating unit 1.
Switching control may be performed according to the V / UV discrimination data from 07. Quantized output data from such a vector quantization unit 110 is extracted via an output terminal 111.

The high-precision pitch search unit 106
, High-precision (fine) pitch data is encoded by the pitch encoding unit 115 and extracted via the output terminal 112. Further, the voiced / unvoiced sound (V / UV) discrimination data from the voiced / unvoiced sound discriminating unit 107 is extracted via an output terminal 113. Each of these output terminals 1
The data from 11 to 113 is transmitted as a signal of a predetermined transmission format.

Each of these data is obtained by processing the data in the block of N samples (for example, 256 samples), and the block is represented on the time axis by the frame of L samples. , The data to be transmitted is obtained in the frame unit. That is, the pitch data, V / UV discrimination data, and amplitude data are updated in the frame cycle.

Next, a synthesizing side (decoding side) for synthesizing an audio signal based on each of the data obtained by transmission.
Will be described with reference to FIG.
15, the input terminal 121 is supplied with the vector-quantized amplitude data, the input terminal 122 is supplied with the encoded pitch data, and the input terminal 123 is supplied with the V / UV discrimination data. You. The quantized amplitude data from the input terminal 121 is sent to the inverse vector quantizer 124 and inversely quantized. The inverse vector quantization unit 124 also has a structure in which the codebook for V (voiced sound) and the codebook for UV (unvoiced sound) are switched and selected according to the V / UV discrimination data obtained from the input terminal 123. ing. Output data from the inverse vector quantization unit 124 is sent to the data number inverse transformation unit 125 and inversely transformed.
The obtained amplitude data is sent to voiced sound synthesis section 126 and unvoiced sound synthesis section 127. The encoded pitch data from the input terminal 122 is decoded by the pitch decoder 128 and sent to the data number inverse converter 125, the voiced sound synthesizer 126, and the unvoiced sound synthesizer 127. V from input terminal 123
The / UV discrimination data is sent to voiced sound synthesis section 126 and unvoiced sound synthesis section 127.

In the voiced sound synthesis unit 126, for example, a cosine (cosin
e) A voiced sound waveform on the time axis is synthesized by wave synthesis, and the unvoiced sound synthesis unit 127 synthesizes an unvoiced sound waveform on the time axis by filtering, for example, white noise with a band-pass filter. The unvoiced sound synthesized waveform is added and synthesized by the adder 129 and extracted from the output terminal 130. In this case, the amplitude data, the pitch data, and the V / UV discrimination data are updated and provided every frame (L samples, for example, 160 samples) at the time of the analysis, but the continuity between the frames is improved (smoothness). For example, each value of the amplitude data and the pitch data is set as a data value at, for example, a center position in one frame, and each data value up to the center position of the next frame (one frame at the time of synthesis) is interpolated. Ask by That is,
In one frame at the time of synthesis (for example, from the center of the analysis frame to the center of the next analysis frame), each data value at the leading sample point and the end point (the leading edge of the next combined frame)
Each data value at a sample point is given, and each data value between these sample points is obtained by interpolation.

Hereinafter, the synthesizing process in the voiced sound synthesizing section 126 will be described in detail. The one synthesized frame (L samples, for example, 160 samples) on the time axis in the m-th band (m-th harmonic band) determined as V (voiced sound)
Assuming that the voiced sound of the minute is V _m (n), V _m (n) = A _m (n) cos (θ _m (n)) 0 using the time index (sample number) n in this synthesized frame. ≦ n <L (12) The final voiced sound V (n) is synthesized by adding (ΣV _m (n)) the voiced sounds of all the bands determined as V (voiced sound) in all the bands.

A _m (n) in the equation (12) is the amplitude of the m-th harmonic interpolated from the top to the end of the synthesized frame. In the simplest case, the value of the m-th harmonic of the amplitude data updated for each frame may be linearly interpolated.
That is, the m-th position at the end (n = 0) of the composite frame
The amplitude value of the harmonic is A _0m , and the end of the synthesized frame (n =
L: the amplitude value of the m-th harmonic at the end of the next synthesized frame is A _Lm , where A _m (n) = (Ln) A _{0 m} / L + nA _Lm / L (13) A _m (n) may be calculated.

Next, the (12) the phase theta _m (n) in the expression by _{θ m (n) = mω O1} n + n 2 m (ω L1 -ω 01) / 2L + φ 0m + Δωn ··· (14) You can ask. In this equation (14), φ _0m indicates the phase of the m-th harmonic (frame initial phase) at the front end (n = 0) of the composite frame, and ω ₀₁ indicates the phase at the front end (n = 0) of the composite frame. The fundamental angular frequency ω _L1 indicates the fundamental angular frequency at the end of the combined frame (n = L: the leading end of the next combined frame). Δω in the above equation (11) is
The minimum Δω is set so that the phase φ _{Lm at} n = L becomes equal to θ _m (L).

The method of obtaining the amplitude A _m (n) and the phase θ _m (n) according to the V / UV discrimination result when n = 0 and n = L in an arbitrary m-th band will be described below. I do. When the m-th band is V (voiced sound) for both n = 0 and n = L, the amplitude A _m (n) is calculated according to the above-described equation (13), and the transmitted amplitude value A _0m , A Amplitude A by linear interpolation of A _Lm
m (n) may be calculated. The phase θ _m (n) is n = 0 and θ
Δω is set so that θ _m (L) becomes φ _Lm when _m (0) = φ _0m and n = L.

Next, when n = 0, V (voiced sound) and n =
If L (unvoiced sound) at L, the amplitude A _m (n)
Is linearly interpolated so that 0 A _m (L) from the transmission amplitude value A _{0 m} of A _m (0). The transmission amplitude value A _Lm at n = L is the amplitude value of the unvoiced sound, and is used in unvoiced sound synthesis described later. Phase θ _m (n) is set to _{θ m (0) = φ 0m} , and [Delta] [omega =
Set to 0.

Further, when n = 0, UV (unvoiced sound)
If V = voiced sound when n = L, the amplitude _Am
(n) sets the amplitude A _m (0) at n = 0 to 0 and performs linear interpolation so that the transmitted amplitude value A _Lm at n = L. Phase θ
The _m (n), the phase theta _m as (0) at n = 0, by using the phase value phi _Lm at the frame _{end, θ m (0) = φ} Lm -m (ω O1 + ω L1) L / 2 (15) and Δω = 0.

When both n = 0 and n = L are V (voiced sound), Δω is set so that θ _m (L) becomes φ _Lm.
The method for setting is described. In the above equation (14), n
= By placing the _{L, θ m (L) =} mω O1 L + L 2 m (ω L1 -ω 01) / 2L + φ 0m + ΔωL = m (ω O1 + ω L1) L / 2 + φ 0m + ΔωL = φ Lm becomes, organize this Then, [Delta] [omega is, Δω = (mod2π ((φ Lm -φ 0m) - and mL becomes _{(ω O1 + ω L1) /} 2) / L ··· (16) this equation (16) in mod2π (x). Is the main value of x
This function returns a value between π and + π. For example, x = 1.3
mod2π (x) = -0.7π when π, mod2 when x = 2.3π
When π (x) = 0.3π, x = -1.3π, mod2π (x) = 0.7
π, and so on.

FIG. 16A shows an example of the spectrum of an audio signal. Bands (harmonics numbers) m of 8, 9, and 10 are set to UV (unvoiced sound), and other bands are set to UV (unvoiced sound). Is V (voiced sound). This V
The time axis signal of the (voiced sound) band is the voiced sound synthesis unit 1
26, and the time axis signal of the UV (unvoiced sound) band is synthesized by the unvoiced sound synthesis unit 127.

Hereinafter, the unvoiced sound synthesizing process in the unvoiced sound synthesizing section 127 will be described. The white noise signal waveform on the time axis from the white noise generating unit 131 is windowed with an appropriate window function (for example, a hamming window) with a predetermined length (for example, 256 samples), and the STFT (short circuit) is performed by the STFT processing unit 132. By performing a term Fourier transform) process, a power spectrum on the frequency axis of white noise as shown in FIG. 16B is obtained. The power spectrum from the STFT processing unit 132 is sent to the band amplitude processing unit 133, and as shown in FIG. 16C, the amplitude | for the UV (unvoiced) band (for example, m = 8, 9, 10) _Am | _UV is multiplied, and the amplitude of the other V (voiced sound) bands is set to zero. The band amplitude processing unit 133 is supplied with the amplitude data, the pitch data, and the V / UV discrimination data. The output from the band amplitude processing unit 133 is sent to the ISTFT processing unit 134, and the phase is converted to a signal on the time axis by performing inverse STFT processing using the phase of the original white noise. The output from the ISTFT processing unit 134 is output to an overlap adding unit 135.
And repeats overlap and addition while weighting appropriately (to restore the original continuous noise waveform) on the time axis to synthesize a continuous time axis waveform. The output signal from the overlap adding unit 135 is sent to the adding unit 129.

As described above, the respective signals of the voiced sound portion and the unvoiced sound portion that have been synthesized in the synthesis units 126 and 127 and returned on the time axis are added by the addition unit 129 at an appropriate fixed mixing ratio. The reproduced audio signal is extracted from the output terminal 130.

As for the configuration on the voice analysis side (encoding side) in FIG. 10 and the configuration on the voice synthesis side (decoding side) in FIG. 15, each part is described in terms of hardware. It can also be realized by a software program using a signal processor) or the like.

The present invention is not limited to the above embodiment. For example, not only audio signals but also audio signals can be used as input signals. The parameters representing the characteristics of the input audio signal (voice signal or acoustic signal) are not limited to the above-mentioned V (voiced sound) / UV (unvoiced sound) discrimination information, but include the pitch value, the strength of the pitch component, and the signal spectrum. Slope, level, etc. can be used. Further, such a characteristic parameter may be replaced by a part of the parameter information originally transmitted according to the encoding scheme, or may be transmitted separately. It can be regarded as a structured codebook if it is transmitted separately.

[0079]

As is apparent from the above description, according to the encoding method of the present invention, an input audio signal is divided into blocks and converted to a frequency axis. The data on the frequency axis as a dimensional vector is obtained, the data on the frequency axis of the M-dimensional vector is divided into a plurality of groups, and a representative value is obtained for each group, thereby reducing the dimension to the S dimension (S <M). This S
The first vector quantization is performed on the data of the dimensional vector, and the first vector quantized output data is inversely quantized to obtain a corresponding S-dimensional code vector. Extend to a vector of dimensions,
A plurality of codebooks corresponding to audio signal states for performing second vector quantization on data representing the relationship between the expanded M-dimensional vector and the original data on the frequency axis of the M-dimensional vector. Since the quantization is performed by switching the plurality of codebooks in accordance with a parameter representing a feature of each block of the audio input signal using the second vector quantizer having the above, while improving sound quality, The partition position information can be transmitted with a small number of bits.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a schematic configuration of an encoding device (encoder) to which an encoding method according to the present invention is applied.

FIG. 2 is a diagram for explaining a codebook formation (training) method.

FIG. 3 is a block diagram illustrating a schematic configuration of a main part of an encoding device (encoder) for describing a method according to another embodiment of the present invention.

FIG. 4 is a diagram for explaining an operation of vector quantization having a hierarchical structure.

FIG. 5 is a diagram for explaining an operation of vector quantization having a hierarchical structure.

FIG. 6 is a diagram for explaining an operation of vector quantization having a hierarchical structure.

FIG. 7 is a diagram for explaining an operation of vector quantization having a hierarchical structure.

FIG. 8 is a diagram for explaining an operation of vector quantization having a hierarchical structure.

FIG. 9 is a diagram for explaining an operation of vector quantization having a hierarchical structure.

FIG. 10 is a functional block diagram showing a schematic configuration of an analysis side (encoding side) of a speech signal synthesis analysis encoding apparatus as a specific example of an apparatus to which the encoding method according to the present invention is applied.

FIG. 11 is a diagram for explaining windowing processing.

FIG. 12 is a diagram for explaining a relationship between a windowing process and a window function.

FIG. 13 is a diagram showing time axis data as an object of orthogonal transform (FFT) processing.

FIG. 14 is a diagram showing spectrum data on a frequency axis, a spectrum envelope (envelope), and a power spectrum of an excitation signal.

FIG. 15 is a functional block diagram showing a schematic configuration of a synthesis side (decode side) of a speech signal synthesis analysis and encoding apparatus as a specific example of an apparatus to which the encoding method according to the present invention is applied.

FIG. 16 is a diagram for explaining unvoiced sound synthesis when synthesizing an audio signal.

[Explanation of symbols]

12 frequency axis conversion processing unit, 13 nonlinear compression unit,
14 inter-block (inter-frame) difference processing unit, 15
Vector quantizer, 15 _Q , 25 _{1Q to} 25 _SQ
Vector _{quantizer, 15} V, ₂₅ 1V ~25 _SV
V (voiced sound) for the code book, ₁₅ U, _{25 1U} ~
Codebook for 25 _SU UV (unvoiced sound) 15 _W , 25 _{1W to} 25 _SW selector switch, 21
Dimension reduction unit 22 _Q S-dimensional vector quantizer, 22 _C S-dimensional codebook, 23-dimensional expansion unit, 103 pitch extraction unit, 104 window processing unit, 105 orthogonal transform (FFT) unit, 106 high-precision (fine) pitch Search section 107 voiced / unvoiced (V / UV) discriminating section 108 amplitude re-evaluation section 109 data number conversion (data rate conversion) section 110 vector quantization section 126 voiced sound synthesis section 127 unvoiced sound synthesis section

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-2-146100 (JP, A) JP-A-2-287500 (JP, A) JP-A-59-12499 (JP, A) JP-A-4- 75100 (JP, A) JP-A-5-265487 (JP, A) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 19/00-19/14 H03M 7/30 H04B 14/04

Claims (1)

(57) [Claims] A step of obtaining data on a frequency axis as an M-dimensional vector based on data obtained by dividing an input audio signal into blocks and converting the data into a frequency axis; Dividing the upper data into a plurality of groups and obtaining a representative value for each group to reduce the dimension to S dimension (S <M); and performing first vector quantization on the data of the S dimension vector. Applying, dequantizing the first vector quantized output data to obtain a corresponding S-dimensional code vector, extending the S-dimensional code vector to an original M-dimensional vector, An audio for performing the second vector quantization on the data representing the relationship between the expanded M-dimensional vector and the original data on the frequency axis of the M-dimensional vector (E) using a second vector quantizer having a plurality of codebooks corresponding to the state of the signal, performing quantization by switching the plurality of codebooks in accordance with parameters representing characteristics of each block of the audio input signal; And a coding method.