KR20040043278A - Speech encoder and speech encoding method thereof - Google Patents

Abstract

PURPOSE: A vocoder and a voice coding method using the vocoder are provided to reduce the quantity of calculations required for fixed codebook search for innovation analysis and decrease the number of memories and bits used for the fixed codebook search. CONSTITUTION: A vocoder includes a pre-processor for pre-processing an input voice signal, a linear predictive coding analysis and quantization unit for carrying out linear predictive coding analysis and quantization for the pre-processed signal, a pitch analyzer for analyzing a pitch to obtain long-section correlation, and a codebook searching unit for searching an algebraic codebook to model an innovation signal. The vocoder further includes a synthesis filter for synthesizing a speech signal using the pitch and code vector searched by the pitch analyzer and the codebook searching unit and a linear predictive coding coefficient output from the linear predictive coding analysis and quantization unit, an adder for calculating an error between the outputs of the pre-processor and the synthesis filter, a weight filter for receiving the error obtained by the adder to output a weight error, and a parameter encoder for encoding the linear predictive coding coefficient, a pitch parameter selected by the pitch analyzer and a codebook parameter selected by the codebook searching unit to output encoded voice data.

Description

Speech Encoder and Speech Encoding Method Using The Same {SPEECH ENCODER AND SPEECH ENCODING METHOD THEREOF}

The present invention relates to a speech signal processing method, and more particularly, to an encoder and a method of a wideband speech signal having a sampling period of 16 kHz.

In general, a target signal for speech signal processing may be divided into a narrowband signal and a wideband signal based on the bandwidth of the signal. Narrow-band signal means using 16-bit linear PCM (linear pulse code modulation) data which sampled the analog input audio signal at 8 kHz as the input signal of the encoder. This means using 16-bit linear PCM data as the input signal of the encoder.

Speech coding techniques (codecs) for narrowband signals are based on the ITU-T (International Telecommunications Union Telecommunication) standard, which is based on the PCM method of G.711 to G.712 and other PCM methods such as G.720 to G.729 series. There is a way to compress. In addition, voice codecs for wideband signals include G.722 and G.722.1 of ITU-T, and AMR-WB (G.722.2), which is to be used for IMT-2000.

ITU-T G.723.1 is one of the representative narrowband voice coders, standardized to compress multimedia signals at low speed, and is an algorithm that compresses and recovers input voice at a dual rate of 5.3 / 6.3 kbit / s. Toll quality. In addition, G.723.1 uses a hybrid coding technology that combines waveform coding and parametric coding, and is a code encoder of CELP (Code Excited Linear Prediction) series.

ITU-T G.722 is standardized for encoding wideband audio signals. It has a data rate of 64,56,48 kbit / s, has face-to-face communication quality, and two sub After dividing into sub-bands, each band is encoded using AD-PCM.

3GPP AMR-WB is the latest standardized wideband voice encoder. It was standardized for application to IMT-2000 to expand the demands of mobile communication. The same codec was named G.722.2 in ITU-T and standardized for simultaneous use in wired and wireless. G.722.2 has nine data rates, up to 23.85 kbit / s, and offers sound quality exceeding ITU-T G.722 64 kbit / s at the maximum data rate.

On the other hand, the recently used CELP speech codec can be divided into LPC analysis (linear predictive coding analysis), pitch analysis (Pitch analysis) and innovation analysis (innovation analysis), in particular to propose a variety of ways to model innovation It is. Among them, a fixed codebook of an algebraic structure having excellent performance is widely used.

The fixed codebook of the algebraic structure is composed of three structures, namely, a method of constructing a codebook structure, an indexing method of selected samples, and a method of searching a codebook. In particular, the algebraic codebooks of narrowband speech coders are often used due to their small amount of computation.

However, since the wideband speech signal has a sampling rate of 16 kHz and has twice as many samples as the narrowband speech signal, the algebraic codebook has many problems in calculation amount and memory usage when applied to the wideband speech coder.

Therefore, the present invention is to solve this problem, fixed codebook structure that can reduce the amount of computation used in the fixed codebook search for innovation analysis, the use memory and the use bit in the CELP wideband speech coding method and using the same An object of the present invention is to provide a fixed codebook retrieval method.

1 is a diagram showing the structure of a typical CELP encoder.

2 is a diagram illustrating the structure of a decoder of a general CELP scheme.

3 is a diagram illustrating the structure of a general AMR-WB 12.65 kbit / s algebraic codebook.

4 is a diagram illustrating a partitioned algebra codebook structure according to an embodiment of the present invention.

*** Description of the symbols for the main parts of the drawings ***

101: preprocessor 102: LPC analysis unit

103: synthesis filter 104: pitch analysis unit

105: fixed codebook search unit 108: fixed codebook

110: adaptive codebook 113: parametric encoder

The algebraic codebook structure according to the characteristics of the present invention for achieving the technical problem is a predetermined number of serial pulse numbers in an algebraic codebook structure in subframe units in a speech encoder for converting an input speech signal into encoded speech data. And a series of tracks corresponding to the respective pulse numbers and pulse positions belonging to the respective track numbers, divided into first and second parts, and the pulse numbers and pulse positions divided into the respective parts.

The pulse position included in each part may be less than or equal to 1/2 of the number of subframe samples.

In addition, the speech encoder according to an aspect of the present invention, a speech encoder for converting the input speech signal into the encoded speech data, the speech encoder comprising: a preprocessor for performing a pre-processing process for the input speech signal; An LPC analysis and quantization unit performing linear predictive coding (LPC) analysis and quantization processing on the preprocessed signal; A pitch analyzer for analyzing pitches to obtain long-term correlations; Algebraic codebook to model the innovation signal excluding the contribution to the long-term correlation in the excitation signal, where the algebraic codebook consists of subframe units, and the structure of the algebraic codebook includes a predetermined number of serial pulse numbers and each pulse number A codebook search unit comprising a series of tracks corresponding to and a pulse position belonging to each track, divided into a first part and a second part, wherein the pulse number and the pulse position are divided into the respective parts. ; A synthesis filter for synthesizing a speech signal using the pitch and code vectors retrieved by the pitch analyzer and the codebook search unit and the LPC coefficients output from the LPC analyzer and the quantization unit; An adder for obtaining an error between the output of the preprocessor and the synthesis filter; A weight filter for outputting a weight error by inputting the error obtained by the adder; And a parameter encoding unit configured to output the encoded speech data by encoding the LPC coefficient, the pitch parameter selected by the pitch analyzer, and the codebook parameter selected by the codebook search unit.

The pulse position included in each part may be less than or equal to 1/2 of the number of subframe samples.

According to another aspect of the present invention, there is provided a speech encoding method comprising: a) performing a preprocessing process on an input speech signal; b) performing linear predictive coding (LPC) analysis and quantization on the preprocessed signal; c) an algebraic codebook for analyzing the pitch and generating a codevector for synthesizing the input speech signal, wherein the algebraic codebook is in subframe units, and the structure of the algebraic codebook is a predetermined number of serial pulse numbers, each pulse A series of tracks corresponding to the number and pulse positions belonging to the respective tracks, each divided into a first part and a second part, wherein the pulse number and the pulse position are divided into the respective parts. The searching of the algebraic codebook may include performing an algebraic code search on a pulse belonging to the first part, searching for an algebraic code that minimizes an error with a target vector for an input speech signal, and then searching for an algebraic codebook belonging to the second part. Algebraic code search for pulses to minimize error with target vector for input speech signal Must navigate; d) synthesizing a speech signal using the pitch and code vectors retrieved by the pitch analysis and codebook search and the LPC coefficients output from the LPC analysis and quantization unit; And e) encoding the LPC coefficients, the pitch parameters selected by the pitch analysis, and the codebook parameters selected by the codebook search to output encoded speech data.

C),

Selecting positions of the pulses sequentially with respect to the pulses belonging to the first part; Selecting positions of the pulses sequentially with respect to the pulses belonging to the second part; And selecting a sign for the pulse at which the position is selected.

Hereinafter, the present invention will be described in detail with reference to the accompanying drawings the most preferred embodiment that can be easily carried out by those of ordinary skill in the art as follows.

1 is a diagram illustrating a configuration of a general CELP speech coder.

As shown in FIG. 1, the preprocessor 101 performs a preprocessing process on the input speech signal s (n), and the linear predictive coding (LPC) analysis and the quantizer 102 are performed by the preprocessor 101. LPC analysis and quantization processing are performed on the preprocessed speech signal and the LPC coefficients are output. In the synthesis filter 103, the LPC coefficients output from the LPC analysis and quantizer 102 are quantized for transmission, and the quantized coefficients constitute the synthesis filter 103. The synthesis filter 103 configured as described above is used as a filter for synthesizing the final speech signal using the results (pitch and code vector) searched in the analysis-by-synthesis process. The function of the synthesis filter 103 is as shown in Equation 1 below.

In the above expression _i (i = 1, ..., m) is the quantized LPC coefficient. In addition, in the above equation, the prediction order is determined according to the value of m. In general, the narrowband speech codec uses the 10th order, and the wideband speech codec uses the order of about 16-20.

After the synthesis filter 103 is configured, the excitation signal is obtained by turning to the closed loop using the constructed synthesis filter. LPC coefficient analysis is performed only once per frame, and the excitation signal is obtained in subframe units. In order to construct a synthesis filter for each subframe, the LPC coefficients obtained in the current frame form a synthesis filter for each subframe through interpolation with the LPC signal obtained in the previous frame.

The excitation signal includes a long term correlation signal obtained from the adaptive codebook 110 and a signal obtained from the fixed codebook 108. These two signals are multiplied by the appropriate gains and then summed to form the excitation signal of the synthesis filter. Equation 2 below represents a function of the synthesis filter of the long-term correlation signal obtained from the adaptive codebook 110.

Where T is the pitch period and g _p is the gain. That is, the current signal is predicted in the long period using the past synthesized signal before the period T, and multiplying the gain by the gain results in the current long-term correlation signal.

After obtaining T and g _p , which are main coefficients of the long-term correlation, the pitch analysis 104 and the codebook search 105 are performed to obtain a more precise excitation signal. In this case, the output of the synthesized signal (output of the synthesis filter 103) is obtained for the pitch period and the gain in all the given cases and all the code vectors in the codebook.

The adder 106 obtains an error between the output of the preprocessor 101 and the output of the synthesis filter 103, and the adaptive weight filter 107 inputs the error obtained by the adder 106 to obtain a weighted error. Output Pitch analysis 104 and codebook search 105 are performed to minimize the energy of the weight error.

After the processes of the LPC analysis 102, the pitch analysis 104, and the codebook search 105 are performed as described above, the parameter encoder 113 encodes the LPC coefficients, the selected pitch parameter, and the codebook parameters by inputting the input voice. The signal s (n) is converted into bit stream data and output.

The target signal of the fixed codebook search 105 is a signal remaining after subtracting the long-term correlation signal from the target signal for obtaining the long-term correlation. Fixed codebooks are implemented in a variety of ways, with the most recent use of the algebraic codebook structure. This method does not require memory to store codebooks and can obtain desired innovation signals at high speed. In addition, this structure has a disadvantage in that a large amount of calculation, and a variety of fast algorithm has recently been proposed to reduce the amount of calculation. However, the fixed codebook search routine still occupies the largest amount of computation in the entire encoding process.

In this method, a criterion for obtaining an optimal pulse is shown in Equation 3 below.

In the above equation, the innovation vector c _k is the number of pulses selected in advance in the codebook.

Further, the vectors d and φ are calculated in advance by the following equations (4) and (5), respectively.

Here, x ₂ (n) is a target signal for fixed codebook search, and h (n) is an impulse response of a weighted synthesis filter.

As described above, the general CELP speech coder as shown in FIG. 1 selects the pitch and code vector when it is the closest to the input signal while actually synthesizing the speech signal for all possible pitch and codebooks. This is necessary. Therefore, in general, the codebook is first ignored and only the pitch is searched. Then, the selected pitch result is fixed and the codebook is searched.

On the other hand, since the size of a fixed codebook is quite large in a general CELP speech coder, a search for a code vector in many cases is required. In addition, since the fixed codebook searches more frequently than the pitch, the calculation amount for the fixed codebook search occupies the largest portion in the calculation amount of the entire CELP speech coder.

Therefore, the structure of the fixed codebook in the CELP speech coder is a very important part in determining the amount of computation and the performance of the speech coder. If the structure of the fixed codebook is very simple and small in size, the computational complexity for the search is reduced, but the performance of the encoder is poor because it cannot display various types of excitation signals. On the other hand, if the size of the fixed codebook is large, the performance of the encoder can be improved, but the amount of computation for searching is significantly increased. Further, even when calculating to find the actual positional information, the amount of calculation increases in accordance with the positional information of the track.

Therefore, in AMR-WB, the use of a depth first search method can reduce the use of memory, and in fact, 16 (track) × 16 (track) × 4 = 1024 (word) memory is required. However, using the full search method requires memory of 16 (track) x 16 (track) x 6 = 1536 (word).

On the other hand, Figure 2 shows a typical CELP decoder.

As shown in FIG. 2, the decoder configures the LPC synthesis filter 206 using the bit stream transmitted from the encoder, decodes the indexes of the fixed codebook 202 and the adaptive codebook 204, and multiplies the gains. Generate an excitation signal. Passing this signal through the LPC synthesis filter 206 generates a synthesized signal, which is made to be audible through the post-processing filter 207.

Referring to the structure of a fixed codebook according to an embodiment of the present invention in detail.

The typical AMR-WB has a sampling period of 16 ms since the wideband speech coder has a bandwidth of 7 ms. The codec uses a split band structure and is divided into lower band and upper band based on 6.4 dB. Among them, the upper band contains little information, so it is simply modeled, and the lower band including the important signal is encoded by the CELP method. Since the frame size of this codec is 20ms, the number of samples of one frame is 256 samples for the lower band. Since we have four subframes here, 64 samples are processed in one subframe.

Figure 3 shows the structure of an algebraic codebook of 12.65 kbit / s in the conventional AMR-WB corresponding to these 64 samples.

As shown in FIG. 3, each track has 16 possible positions, pulses are assigned to each track, and each pulse may only exist within the position specified in each track. In addition, the number of pulses determined in each track depends on the transmission rate.

In other words, when a position is determined in each track, a sign of the corresponding pulse is determined. At this time, one sine is determined for each pulse, and if only the sine information and the position information are provided, a pulse obtained from the fixed codebook can be generated.

On the other hand, the performance of the fixed codebook generally determines the configuration of the codebook structure, pulse indexing (pulse indexing) method, codebook search method. The present invention is to configure the structure of the codebook of the split (split) structure to have many advantages in pulse indexing and codebook search, and to reduce the amount of memory used.

4 illustrates a structure of an algebraic codebook having a split structure according to an embodiment of the present invention.

As shown in FIG. 4, an algebraic codebook according to an embodiment of the present invention divides a codebook configured in one subframe unit into two parts having smaller sizes. That is, when the codebook structure according to the embodiment of the present invention is compared with the general codebook structure of FIG. 3, it can be seen that the track size is reduced.

In addition, according to the codebook structure according to the embodiment of the present invention, all search processes can be independently performed with only the first half of the codebook. In other words, in order to obtain position and sign information, necessary information is calculated in advance through Equation 3 and stored in a memory. The algebraic codebook of a split structure according to an embodiment of the present invention is a fixed codebook for one part. When performing a search, only the information about the half of the location is calculated and stored in the memory, and then the search is performed. In addition, since the search is performed for each part in the position search, the search can be performed with much less computation amount than the search for the same position.

For example, in the conventional codebook structure, if each track has 16 positions, 16 x 16 = 256 calculations are needed to find all positions for the two tracks. However, according to an embodiment of the present invention, if a track having 16 positions is divided into two parts consisting of eight positions, the number of positions that can be navigated to the positions of each part is 8 × 8 = 64 times, which is two As part of the process, 64 × 2 = 128 searches. Thus, the search position is reduced by almost half.

Also, even in pulse indexing, since the position of each track is small, position information can be encoded with a few bits.

Meanwhile, in the algebraic codebook of the split structure according to the embodiment of the present invention, n1 and n2 values are important. In FIG. 4, 0,... Corresponding to the positions of pulses K * index ₀ , K * index ₁ , K * index ₂ , and K * index ₃ . , n1 can be recombined in various ways and put in a proper track, but if a value larger than n1 is included, the search process cannot be performed independently.

For example, applying the structure of G.722.2 to the algebraic codebook of the split structure according to the embodiment of the present invention, n1 becomes 31 and n2 becomes 63 in FIG. This limits the position of the pulse to half, which can result in some degradation in sound quality.

However, when the algebraic codebook structure configuration method according to the embodiment of the present invention is used, all processes except for LPC analysis and quantization are performed in subframe units. Therefore, it is possible to save bits when performing pulse indexing, search at a high speed during codebook search, and reduce memory usage.

In addition, the number of pulses allocated to each track and the sign information of each pulse may be variously set by a designer.

The drawings and detailed description of the invention are merely exemplary of the invention, which are used for the purpose of illustrating the invention only and are not intended to limit the scope of the invention as defined in the appended claims or claims. Therefore, those skilled in the art will understand that various modifications and equivalent other embodiments are possible from this. Therefore, the true technical protection scope of the present invention will be defined by the technical spirit of the appended claims.

When the codebook of the split structure according to the present invention is applied as described above, less position information is provided for each sub-part for each search, and memory usage is reduced. Accordingly, since the position to be searched is reduced, the amount of calculation is also reduced, and fixed codebook search can be performed at high speed.

In addition, since the position information is small, a lot of bits can be saved even during the indexing process of the pulse. In addition, since the codec can be reduced, it can be implemented with a small processor when applied to the system, and the cost of the entire system can be reduced.

Claims (6)

In the algebraic codebook structure of a subframe unit in a speech encoder for converting an input speech signal into encoded speech data, A predetermined number of serial pulse numbers, a series of tracks corresponding to each pulse number, and pulse positions belonging to each track, Divided into a first part and a second part, wherein the pulse number and the pulse position are divided into the respective parts, Algebraic codebook structure. The method of claim 1, The pulse position included in each part is less than or equal to 1/2 of the number of subframe samples. Algebraic codebook structure. In the speech encoder for converting the input speech signal to the encoded speech data, A preprocessing unit performing a preprocessing process on the input voice signal; An LPC analysis and quantization unit performing linear predictive coding (LPC) analysis and quantization processing on the preprocessed signal; A pitch analyzer for analyzing pitches to obtain long-term correlations; Algebraic codebooks for modeling innovation signals excluding contributions to long-term correlations in an excitation signal, wherein the algebraic codebooks are composed of subframe units, and the structure of the algebraic codebooks includes a predetermined number of serial pulse numbers A codebook search unit comprising a series of tracks corresponding to and a pulse position belonging to each track, divided into a first part and a second part, wherein the pulse number and the pulse position are divided into the respective parts. ; A synthesis filter for synthesizing a speech signal using the pitch and code vectors retrieved by the pitch analyzer and the codebook search unit and the LPC coefficients output from the LPC analyzer and the quantization unit; An adder for obtaining an error between the output of the preprocessor and the synthesis filter; A weight filter for outputting a weight error by inputting the error obtained by the adder; And A parameter encoder which outputs encoded speech data by encoding the LPC coefficient, the pitch parameter selected by the pitch analyzer, and the codebook parameter selected by the codebook searcher Speech signal encoder comprising a. The method of claim 3, The pulse position included in each part is less than or equal to 1/2 of the number of subframe samples. Speech signal encoder. In the speech encoding method for converting an input speech signal into encoded speech data, a) performing a preprocessing process on the input voice signal; b) performing linear predictive coding (LPC) analysis and quantization on the preprocessed signal; c) an algebraic codebook for analyzing the pitch and generating a codevector for synthesizing the input speech signal, wherein the algebraic codebook is in subframe units, and the structure of the algebraic codebook is a predetermined number of serial pulse numbers, each pulse A series of tracks corresponding to the number and pulse positions belonging to the respective tracks, each divided into a first part and a second part, wherein the pulse number and the pulse position are divided into the respective parts. The searching of the algebraic codebook may include performing an algebraic code search on a pulse belonging to the first part, searching for an algebraic code that minimizes an error with a target vector for an input speech signal, and then searching for an algebraic codebook belonging to the second part. Algebraic code search for pulses to minimize error with target vector for input speech signal Must navigate; d) synthesizing a speech signal using the pitch and code vectors retrieved by the pitch analysis and codebook search and the LPC coefficients output from the LPC analysis and quantization unit; And e) outputting encoded speech data by encoding the LPC coefficient, the pitch parameter selected by the pitch analysis, and the codebook parameter selected by the codebook search. Speech signal encoding method comprising a. The method of claim 5, C), Selecting positions of the pulses sequentially with respect to the pulses belonging to the first part; Selecting positions of the pulses sequentially with respect to the pulses belonging to the second part; And Selecting a sign for the pulse at which the position is selected Speech signal encoding method comprising a.