JPH09297597A - High-efficiency speech transmission system and high-efficiency speech transmission device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce a hardware scale and to enable high-efficiency speech transmission of a low bit rate in a wide band by executing encoding processing to the input speech signal of a past frame by using the speech information parameter formed by subjecting the input speech signal of the past frame to a linear prediction analysis. SOLUTION: The encoding processing to the input speech signal of the N-th frame is executed by using the speech information parameter formed by subjecting the input speech signal of the past frames including the (N-1) the frame to the linear prediction analysis before the execution of the high-efficiency encoding to the input speech signal of the N-th frame. In such a case, speech information parameters, such as filter coefft. and pitch frequency, are obtainable from first, second adaptation devices 22, 32 and a pitch period extraction part 31. The short-period prediction filter coefft. is formed by subjecting a synthesized signal which is the output of a short period prediction filter 42 to the linear prediction analysis with this device and, therefore, an encoder is no more required to transfer the information of the short period prediction filter coefft. to a decoder side.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a high-efficiency voice transmission method and a high-efficiency voice transmission device used in, for example, multimedia communication.

In recent years, with the development of ISDN networks and multimedia information devices, multimedia communication of images, voices, documents, etc. has become more active. Among multimedia communications, in particular, video conferencing systems and visual telephone systems require wideband voice (50 Hz to 7 KHz band) that is more natural and higher in quality than telephone band voice. This is because the wideband sound improves the naturalness and the realism of the sound.

On the other hand, conventional 16KHz sampling, 7KHz
The band audio coding system is standardized as ITU-T G.722, but since the bit rate is 64 to 48 Kb / s, only one channel of audio signal can be transmitted on the line of 64 KHz transmission band.

Therefore, by lowering the bit rate of the audio signal and narrowing the transmission band, other media information can be transmitted in the above-mentioned transmission band and the communication cost can be reduced. Correspondingly, it is necessary to provide a high-efficiency voice transmission method and a high-efficiency voice transmission device with a reduced hardware scale and a wide band and low bit rate.

[0005]

2. Description of the Related Art FIG. 21 is an example of a main part configuration diagram of a voice signal generation model, FIG. 22 is an explanatory diagram of vectors / frames, and FIG. 23 is an explanatory diagram of processing timing of a conventional encoder.

First, the above-mentioned ITU-T G.722 voice encoding system uses a band-division ADPCM system. Therefore, when processing a voice signal, the 7 KHz band is set to 3.4 KHz or higher (high band side).
Divide into 4KHz or less (low frequency side) and divide the audio signal into 16KH
Encode each with a sampling period of z.

At this time, the high band side is coded with 2 bits and the low band side is coded with 6 bits, multiplexed and transmitted to the line at an 8 KHz cycle, resulting in a bit rate of 64 Kb / s as a whole, and 64 KH / s.
The z-band line can only transmit 1ch audio signals.

On the other hand, in the communication in the telephone band (300 to 3.4 KHz band), there is the international standard G.728 (LD-CELP), and the required transmission rate is as low as 16 Kbps. For example, in a video conferencing system, images and audio coexist within a limited transmission rate, so the expansion of the audio communication transmission rate affects the image communication quality. On the contrary, if the image quality is selected to be high, the naturalness of the sound is lost.

In this CELP system, a plurality of voice samples are combined into one group consisting of a vector and a frame, and an error between a voice signal synthesized by using a voice signal generation model described later and an actual input voice signal is minimized. In this method, the voice information parameter is changed so that the minimum parameter is quantized and transmitted.

Here, the voice signal generation model will be described. As shown in Fig. 22, a sample is a collection of multiple samples of the input audio signal and the PCM-converted audio signal (hereinafter referred to as the input audio signal) is called a vector. There are n vectors (n is a positive integer). ), The collection is called a frame.

In FIG. 21, an input voice signal for one frame is temporarily (for example, 30-40 ms) stored in the buffer 81 and then added to the linear prediction analysis part (LPC) 85 and the delay part 82.

The linear predictive analysis portion 85 performs linear predictive analysis on the input audio signal for one frame to extract audio information parameters including a filter coefficient and a gain value, and sends them to the audio synthesis filter 88 and the amplifier 87. To do.

The voice information parameter output from the linear prediction analysis section 85 is updated every time T _0, but is fixed during that period. Further, the vectorization section 83 extracts the input speech vector to be encoded (for example, the input speech vector of # 0 in FIG. 22) from one frame of the input speech signal input through the delay section 82. And add it to the comparison part 84.

On the other hand, a vector quantization codebook (hereinafter referred to as VQ codebook) 86 stores Q kinds of noise signal patterns (hereinafter referred to as sound source patterns) having a length of one vector.

Therefore, for example, the sound source pattern of # 0 extracted from the codebook is passed through the amplifier 87 and the voice synthesis filter 88 having the above filter coefficient to synthesize the voice vector, and the synthesized voice vector is sent to the comparison portion 84. Add.

The comparison section 84 compares the synthesized speech vector with the actual input speech vector to obtain an error.
Repeat (Q-1) sound source pattern Q times to quantize the gain value and filter coefficient when the error is minimum.

If there are 1024 kinds of sound source patterns (Q) in the codebook, the pattern can be specified.
Requires 10 bits. Also, 1 vector is 5 samples,
If the sample period is 8 KHz, the bit rate will be [10 bits / (5 × 125 μs) = 16 Kbs].

However, one vector has n = 10 samples,
If the number of samples is doubled and the same calculation as above is performed, it becomes 8 Kbs. That is, if the number of samples per vector is increased, the bit allocation per sample can be reduced, and the bit rate can be reduced.

Next, the processing timing of the encoder will be described with reference to FIG. 23 (see FIG. 21). As shown in FIG. 23, for example, in order to perform an encoding process on an #N frame input speech signal, it is necessary to perform a linear prediction analysis (hereinafter, referred to as LPC analysis) before this process. The LPC analysis requires the input audio signals up to several samples before and after the input audio signal of the #N frame to be encoded (the shaded area in Fig. 23).
, This part is stored in the buffer 81.

Therefore, the LPC analysis processing is performed from the first half to the second half of the # (N + 1) frame, and the encoding processing is # (N +).
1) It will be performed from the latter half of the frame to the latter half of # (N + 2).

Therefore, the voice coded information obtained by coding the voice information parameter for the #N frame is transmitted at the time of the # (N + 2) frame, and eventually the coded information is obtained with a delay of about 3 frames.

[0022]

As described above, since the LPC analysis is performed on the input voice signal of the frame to be coded from now on, the memory for storing the input voice signals of three frames necessary for the LPC analysis is used. (The buffer in Figure 21) is required. This causes a problem that the hardware scale of the encoding unit becomes large.

Further, for example, in a video conference system, a reproduced audio signal is a wide band and a low bit rate audio encoding method in order to obtain a naturalness and a sense of presence in the audio communication without affecting the quality of image communication. There was a problem that I had to provide.

It is an object of the present invention to provide a high-efficiency voice transmission method and a high-efficiency voice transmission device which has a wide bandwidth and a low bit rate in a reduced hardware scale.

[0025]

FIG. 1 is a functional explanatory view of a main part of a speech coder of the first to third inventions, and FIG. 2 is a functional explanatory view of a main part of a speech coder of the fourth invention. FIG. 3 is an explanatory diagram of the processing timing of FIG. 1, and FIG. 4 is an explanatory diagram of the LPC analysis window.

FIG. 5 shows a speech coder according to the seventh to thirteenth aspects of the present invention.
It is a principal part functional explanatory view of a decoder. The first aspect of the present invention, prior to performing high-efficiency coding on the input speech signal of the Nth frame, performs linear prediction on the input speech signal of the past frame including the (N-1) th frame. Using the voice information parameter generated by the analysis, the input voice signal of the Nth frame is encoded.

According to the second aspect of the present invention, a linear prediction analysis window for weighting an input speech signal in a frame closer to the Nth frame among the past frames is used.

A third aspect of the present invention stores voice information generating means for performing a linear predictive analysis of an input voice signal to generate voice information parameters, and a plurality of noise signal patterns serving as a sound source, and the extracted noise. Vector quantization codebook means for giving a gain to a signal pattern and outputting it, a long-term prediction filter using applied voice information parameters as filter coefficients, a short-term filter, and a synthesized voice signal using the output of the vector quantization codebook means. In a high-efficiency speech coder having a speech synthesis filter means for generating a noise and a minimum error square calculation means for repeating a calculation of an error between an input speech signal and a generated synthesized speech signal to determine a sound source number having a minimum error. , The (N-1) th frame is generated by performing the linear prediction analysis on the input speech signals of the past frames. Using voice information parameters, and configured to perform a coding process for the input audio signal of the N-th frame.

The fourth aspect of the present invention is configured to use, as a filter coefficient of the above-mentioned short-term prediction filter, a voice information parameter generated by performing a linear prediction analysis on a synthesized voice signal output in the past.

In a fifth aspect of the present invention, the input speech coded information is separated to separate the sound source number and the sound information parameter, and the noise signal pattern corresponding to the applied sound source number is extracted to give a gain. In the high-efficiency speech decoder including the vector quantization codebook means for outputting, the long-term prediction filter having the filter characteristics corresponding to the applied speech information parameter, and the short-term prediction filter, the short-term prediction filter has been output in the past. The speech information parameter generated by performing the linear prediction analysis on the decoded speech signal is used as the filter coefficient.

A sixth aspect of the present invention uses the high-efficiency speech encoder of claim 4 to encode and transmit an input speech signal, and encodes the speech using the high-efficiency speech decoder of claim 5. The signal is received and decoded.

In a seventh aspect of the present invention, when a convolution operation of an impulse response matrix and a code vector is performed by a synthesis filter, the operation sequence is started from the most recent code vector, and the past code vector is sequentially operated. Then, the convolution operation is performed.

According to the eighth aspect of the present invention, when the pitch search of the adaptive codebook is performed in units of m vectors forming a frame, the first vector searches the entire range to find the optimum pitch P ₁
After calculating, from the second vector to the m-th vector,
With the optimum pitch found in the previous pitch search as the center, the pitch search is performed within a preset range to find the optimum pitch.

The ninth aspect of the present invention is a convolution operation of an addition code vector obtained by adding code vectors respectively taken out from the adaptive codebook means and the fixed codebook means and an impulse response vector from the impulse response vector generating means. Is provided to provide an impulse response filter means for obtaining a 0-state response vector, and a combination of optimum code vectors that minimizes the error between the 0-state response vector from the impulse response filter means and the input target vector is searched. In the high-efficiency speech coding method, the convolution operation is performed by using the seventh aspect of the present invention.
A pitch search is performed using the present invention.

The tenth aspect of the present invention reproduces the original voice signal by decoding the signal encoded by the voice encoding method of the ninth aspect of the present invention. The 11th aspect of the present invention is configured to perform the convolution operation using the 7th aspect of the present invention and the pitch search using the 8th aspect of the present invention.

The twelfth aspect of the present invention has a configuration in which the signal encoded by the high-efficiency voice encoder of the tenth aspect of the present invention is decoded to reproduce the original voice signal. A thirteenth aspect of the present invention encodes an input voice signal by using the high-efficiency speech encoder of the eleventh aspect of the invention and transmits the speech signal, and encodes the speech signal using the high-efficiency speech decoder of claim 12. Is received and decrypted.

First, the means of the present invention will be described with reference to FIGS. In FIG. 1, as explained in the conventional example, the VQ codebook 51 is prepared in advance with Q kinds of sound source patterns having a length of one vector.

The sound source pattern output from the VQ codebook 51 is provided with a gain corresponding to the output from the gain adaptor 53 in the gain portion 52, and a synthesis filter (composed of a long-term prediction filter and a short-term prediction filter) for synthesizing speech is provided. ) 4, a synthesized speech signal is obtained, but this synthesized speech signal is generated only for the types of sound source patterns prepared in the VQ codebook.

The loop composed of the VQ codebook means 5, the synthesis filter 4, and the minimum error square calculation means 6 is a portion for calculating the error between the above-mentioned synthesized voice signal and the input voice signal, and when the error becomes the minimum. The VQ codebook sound source pattern number (hereinafter referred to as the index) is calculated.

Here, the sound source pattern prepared in the VQ codebook 51 corresponds to the vibration pattern of the human vocal cords and the movement of the lips, tongue, etc. at the time of consonant vocalization. It should be noted that, while the vibrations of the vocal cords produce various reverberations depending on the shape of the palate to produce a voice, the synthetic voice filter corresponding to the shape of the palate of a person changes the value of the filter coefficient. The movement is equivalent to a change in the shape of the palate.

That is, the speech coding information is the index selected from the VQ codebook and the filter coefficient of the synthesis filter, and this filter coefficient represents the input speech signal.
It is generated by LPC analysis.

Here, the signal flow in FIG. 1 will be described. The input audio signal is temporarily stored in the vector buffer 12. Therefore, the LPC analysis section 11 obtains the reflection coefficient and the LPC coefficient using the input speech signal stored in the vector buffer, the former as the first adaptor 22, and the latter as the pitch period extraction section 31 and the auditory weighting filter 62. Send to.

The first adaptor 22 is a reflection coefficient codebook.
21 is accessed to take out the filter coefficient closest to the input reflection coefficient and send it to the short-term prediction filter 42 in the speech synthesis filter.

It is assumed that the reflection coefficient codebook 21 stores a table showing the relationship with the filter coefficients corresponding to various input reflection coefficients. In addition, the pitch period extraction unit 31 obtains the LPC prediction residual and the pitch period from the input LPC coefficient and the input speech signal, and outputs the second adaptive unit 32.
To send to.

The second adaptor 32 accesses the long-term prediction coefficient codebook 33 in the same manner as described above, extracts the corresponding filter coefficient using the input LPC prediction residual and pitch cycle, and outputs the speech with the pitch cycle. It is sent to the long-term prediction filter 41 in the synthesis filter.

In the long-term prediction coefficient codebook 33,
It is assumed that a table showing the relationship between LPC prediction residuals and filter coefficients corresponding to various combinations of pitch periods is stored.

Further, the auditory weighting filter portion 62 is provided with a masking characteristic corresponding to the input LPC coefficient. That is, as described above, the voice information parameters such as the filter coefficient and the pitch period can be obtained from the first adaptor 22, the second adaptor 32 and the pitch period extracting portion 31.

On the other hand, the sound source pattern # 0 (having a length of one vector) is extracted from the VQ codebook 51. Therefore,
The gain portion 52 gives a gain corresponding to the output of the gain adaptor 53 to the extracted # 0 sound source pattern, and sends it to the speech synthesis filter 4 including the long-term prediction filter 41 and the short-term prediction filter 42.

As a result, the sound source pattern passes through the speech synthesis filter 4 to generate a synthesized speech signal of one vector,
It is added to the error calculation part 61. Since the corresponding one-vector input audio signal of the input audio signals stored in the vector buffer 12 is also added to the error calculation portion 61, the error between the one-vector synthesized audio signal and the input audio signal is obtained. Is calculated, and the calculation result is added to the minimum error square calculation part 63 via the auditory weighting filter 62.

Therefore, the minimum error square calculation part 63 performs the minimum error square calculation, and this process is repeated Q times to obtain the index of the VQ codebook 51 that minimizes the error power. Then, the index of the VQ codebook, the index of the long-term prediction coefficient codebook, the pitch period, and the index of the reflection coefficient codebook obtained above are sent to the decoder side as audio information parameters.

Here, in the case of FIG. 1, the coefficients of the short-term prediction filter are generated using the input speech signal, and the generated coefficients of the short-term prediction filter are sent to the decoder side. However, in the case of FIG. 2, the coefficients of the short-term prediction filter are generated using the synthesized speech signal, but the coefficients of the short-term prediction filter are not sent to the decoder side, and the decoder side uses the decoded speech signal to perform the short-term prediction filter. The coefficient of is generated.

As a result, the amount of information to be transmitted to the decoder side decreases, and the transmission speed further decreases. Next, the speech signal has a correlation between samples, and it is possible to predict speech waveforms from the past to the future in a short period (several ms).

Therefore, in the present invention, as shown in FIG. 3, before performing the encoding process on the input audio signal of the #N frame,
As the input audio signal necessary for LPC analysis, # (N
-1) Use the input audio signal before the frame.

When performing LPC analysis, an LPC analysis window having the characteristics shown in FIG. 4 is used. The characteristic of this analysis window is the same as the response waveform that appears when an impulse is applied to a filter having a certain transfer function, and is expressed by the following equation.

H (Z) = 1 / (1−αZ ⁻¹ ) ² where 0 <α <1. In other words, in the LPC analysis window,
By giving the characteristics as shown in, the waveforms close to the next frame are weighted so that they are not affected by the past.

As a result, the LPC analysis can be stably performed. Now, in FIG. 3, the input voice signal before the # (N−1) frame is subjected to LPC analysis in the # (N−1) frame to obtain necessary voice information parameters.

Therefore, when the already obtained # (N-1) voice information parameters are applied to the input voice signal of the #N frame to perform the encoding process, the voice encoding is performed at the time of the # (N + 1) frame. Since the information is extracted, the information delay in encoding is one frame.

As a result, it is possible to provide an encoder with a small hardware scale in response to the reduction in memory capacity.
Next, the means of the seventh to twelfth aspects of the present invention will be described with reference to FIG.

There are various techniques for speech coding methods. Broadly speaking, a waveform coding method for the purpose of faithfully representing a waveform and a speech generation model are analyzed and converted into parameters. There is an analysis and synthesis method that synthesizes from the parameters.

The above-mentioned ITU-T G.722 is a kind of waveform coding method. The characteristic of the waveform coding method is that the weight of the information amount per bit is quite large. It is known that the ratio of losing the information that they have is large, and it is not suitable for lowering the bit rate.

On the other hand, the analysis-synthesis method is known to have a high quality up to a relatively low bit rate because the aggregated data is sent as one block to some extent. CELP (Code Excited Line)
ar Prediction) method.

In general, the CELP system is a system effective as a low bit rate coding system in the telephone band, and is described in G.728 (1).
6 Kbps), as well as VS, which is the Japanese standard method for mobile communication.
ELP (Vector Sum Excited Linear Prediction: 6.7 Kbp
s) and PSI-CELP (Pitch Synchronous Inovation-CELP:
It has been realized as an extremely low-rate encoding method such as 3.45 Kbps.

The CELP system will be outlined. The characteristic parameters of the input speech signal are extracted from the input speech signal by linear prediction analysis to construct a synthesis filter. In advance, multiple excitation signals (prepared as a codebook in the sound source) are sequentially converted into a voice signal level by a synthesis filter, the error with the input voice signal is evaluated, and the vector of the error minimum vector is calculated. High-efficiency coding in which the index is transmitted as a code.

The present invention proposes a configuration necessary for handling a voice signal wider than the telephone band by utilizing the high-efficiency information compressibility of the CELP system. As a related invention, "high-efficiency voice transmission method" is proposed. And a high-efficiency voice transmission device ”have been filed by the present applicant, and the present invention is an improvement thereof. This will be described below with reference to FIG.

In the figure, a vectoring means 101 is a part for vectorizing an input speech signal coded by PCM, and a linear prediction analysis means 103 carries out a linear prediction analysis processing of an input speech signal sequence at a certain cycle. This is a part for calculating the reflection coefficient at that time.

The filter coefficient calculation means 104 converts the reflection coefficient calculated by the linear prediction analysis means 103 into the reflection coefficient table 10
After being quantized by 5, the filter coefficient used in each of the synthesis filter 106 and the perceptual weighting filter 107 is calculated.

The adaptive codebook 109 stores the past gain-adjusted excitation signal, that is, the input signal of the synthesis filter determined to be optimum in the past, and the gain means 111 is a part for adjusting the gain.

The fixed codebook 112 stores the driving sound source, and the gain means 114 is a part for adjusting the gain. The adding means 115 is a part that adds the vectors output from the gain means 111 and the gain means 114.

The difference means 102 is a part for obtaining the difference between the input vector and the vector synthesized by the synthesis filter 106. The synthesis filter means 106 is a part for synthesizing a signal from the excitation signal whose gain has been adjusted.

The perceptual weighting filter 107 is a part for perceptually weighting the difference vector between the input vector and the combined vector. The minimum error calculation means 108 is a part that evaluates the weighted error vector.

Note that each unit of the decoder in FIG. 5 performs the same processing as the unit of the corresponding encoder, and therefore its description is omitted. The flow of signals in FIG. 5 is as follows.

The PCM-encoded input voice signal is put into a vector by the vectorizing means 101 in units of n, and this vector is called a target vector. Further, the linear prediction analysis unit 103 performs a linear prediction analysis on the above-mentioned input speech signal for each frame and sends the analysis result to the filter coefficient calculation unit 104. Therefore, the filter coefficient calculation means 104 calculates the auditory weighting filter coefficient and the synthesis filter coefficient using the analysis result and gives them to the corresponding filter.

In calculating the coefficients, a reflection coefficient table prepared by using L-level reflection coefficients of 1 to L obtained by the linear prediction analysis at respective required levels
Quantize at 105, and use the quantized value to convert into filter coefficients.

For each output vector of the adaptive codebook 109, a gain-giving adaptive code vector to which an optimum gain is given by the gain means 111, and for each vector of the fixed codebook 112, an optimum gain is given by the gain means 114. The adding means 115 adds the fixed gain-applying code vectors given by.

Then, the addition result of the addition means 115 is used as the synthesis filter 10 whose filter coefficient is updated for each frame.
The vector of weighting regions ((
Hereinafter, it will be referred to as a weighting signal) and added to the difference means 102.

Since the output from the vectorization means 101 (hereinafter referred to as the target vector) is also added to the difference means 102, the error between the weighting signal and the target vector is calculated.

In calculating the error, the minimum error calculating means 108 evaluates the error after passing through the auditory weighting filter 107 for reflecting the human auditory characteristics. As a result of the error evaluation, the combination of the adaptive code vector and the fixed code vector that minimizes the error is determined, and the adaptive codebook index and fixed codebook index, which are the indices of each code vector at that time, and each codebook gain table. The adaptive codebook gain index and fixed codebook gain index, which are indices 110 and 113, are transmitted to the demodulator as encoded information.

Further, the quantization table index of the reflection coefficient obtained from the reflection coefficient table 105 is transmitted to the decoder as coding information by the linear prediction analysis for each frame period.

Now, in the decoder, each index of the reflection coefficient transmitted in each frame period and the adaptive / fixed codebook index / transmitted in each vector
And adaptive / fixed codebook gain index
Separate the /.

Then, for each frame period, the quantized reflection coefficient corresponding to the reflection coefficient table index is read from the reflection coefficient table 121, and the filter coefficient calculation means 12
Generates and updates the synthesis filter coefficient with 0.

Further, the quantization gain corresponding to the index is read from the adaptive / fixed codebooks 123 and 126 and the adaptive / fixed codebook gain tables 125 and 128 for each vector, and the gain adjustment means 124 and 127 perform gain adjustment. The audio signal is reproduced by passing the signal through the synthesis filter 122 similarly to the encoder.

Here, when a wideband audio signal is handled, the number of formants contained therein increases, so the order L of the synthesis filter for reproducing the frequency characteristic is set to 16 to 20. In order to realize a low rate, both the number of samples per vector and the number of vectors per frame are increased.

In the case of the present invention, the number of samples n per vector is suitably 30 to 40, and the number of vectors m per frame is about 8 to 10. In addition, the CELP method generally has a large amount of calculation, and many methods of reducing the amount of calculation have been proposed.

The present invention provides an arithmetic means for reducing the number of times of the convolution operation of the impulse response matrix of the synthesis filter and the adaptive code vector, which has a particularly large arithmetic amount.

[0085]

FIG. 6 is a functional explanatory diagram (encoder) of the first to third embodiments of the present invention, FIG. 7 is an explanatory diagram of a processing procedure of FIG. 6, and FIG. 8 is an encoder of FIG. FIG. 9 is a functional explanatory diagram of the corresponding decoder, and FIG. 9 is a functional explanatory diagram (decoder) of the sixth embodiment of the present invention.

FIG. 10 is a functional explanatory view (encoder) of the seventh to ninth and eleventh embodiments of the present invention, FIG. 11 is an explanatory view of the processing procedure of FIG. 10, and FIG. 12 is a tenth and twelfth book. Functional explanatory view of the embodiment of the invention (
Decoder) FIG. 13 is an explanatory diagram of the processing procedure of FIG.

FIG. 14 is a diagram for explaining the data storage order. (A) is a normal order storage, in the case of the conventional example, (b) is a reverse order storage. In the case of the seventh embodiment of the present invention, FIG. 14 (a), the calculation amount explanatory diagram (1), FIG. 16 is the calculation amount explanatory diagram (2) in FIG. 14 (a), FIG. 17 is the calculation amount of the seventh embodiment of the present invention FIG. 18 is an explanatory diagram (No. 1) and FIG. 18 is an explanatory diagram (No. 2) of the amount of calculation in the embodiment of the seventh invention.

FIG. 19 is an explanatory diagram of the pitch preliminary search method of the adaptive codebook (when pitch extraction evaluation is performed every time with all vectors), (a) is a search range explanatory diagram, and (b) and (c) are processing procedure explanations. FIG. 20 is an explanatory diagram of the eighth embodiment of the present invention (when only the range of several samples before and after the pitch information of the front vector is evaluated), and (a) is an explanatory diagram of the search method of the present invention. , (B), (c) are processing procedure explanatory diagrams.

Here, the same reference numerals denote the same objects throughout the drawings. Further, the parts described in detail above will be briefly described, and the parts of the present invention will be described in detail. 6 to 9 and 10 to 18 will be described below.

First, the description of FIGS. 6 to 9 will be given. In FIG. 6, the vector buffer portion 12 is an input audio signal (for example, 16-bit linear PCM sampled at 16 KHz).
Sampled audio data) to m samples (m is a positive integer)
Of continuous speech vector signals S (n) are generated.

The first perceptual weighting filter 62a generates a perceptually weighted voice vector V (n) for the above voice vector signal S (n). The long-term prediction filter 41 is a comb filter that raises the level of frequency components that are K times (K is a positive integer) the fundamental pitch frequency of speech, and the short-term prediction filter 42 has a frequency characteristic that envelops the speech spectrum. Is.

Then, the long-term prediction filter 41, the short-term prediction filter 42, and the second auditory weighting filter 62b are connected in cascade, and 0,0,
.. The output of the second auditory weighting filter 62b when 0 is added is calculated (this output is referred to as 0 input response signal).

Now, the first difference between the input sound signal v (n), which is perceptually weighted by the first perceptual weighting filter 62a, and the 0 input response signal r (n) from the second perceptual weighting filter 62b. The calculation part 61a calculates v (n) -r (n) to obtain a difference output (hereinafter, referred to as a target vector) x (n).

A synthetic speech signal Xj generated from the target vector x (n) and the sound source pattern of the VQ codebook 51.
The index of VQ codebook 51 is determined so that the error between and is minimum.

The VQ codebook 51 is prepared in advance with Q kinds of sound source patterns as described above, and the index search of the VQ codebook is performed by the following procedure. 1 from the jth index in VQ codebook 51
A vector sound source pattern (hereinafter referred to as a code vector) is defined as yj (yj = yj (0), yj (1), ..., Yj (n-1)).

The code vector yj is filtered in the gain section 52 after being given a gain σ corresponding to the predicted amplitude of the input signal, and then the output of the impulse response vector 43 is applied to the applied impulse response filter 44. It

Here, the impulse response vector part 43
Is the perceptual weighting filter coefficient, short-term prediction filter coefficient, long-term prediction filter coefficient, which is obtained from the LPC analysis results,
Calculate the impulse response vector using the pitch period,
This response vector is applied to the impulse response filter 44.

Therefore, the impulse response filter 44 has the same response characteristic as the impulse response characteristic when the impulse is input to the long-term prediction filter 41, the short-term prediction filter 42, and the auditory weighting filter 63b connected in cascade. The response signal for one vector of time is h (0), h (1),
.., h (n-1).

As a result, the impulse response filter 44
Is an FIR with filter coefficients h (0), h (1), ..., h (n−1)
Represented by a shape filter.

Now, in the impulse response filter 44,
Let xj (yj = xj (0), xj (1), ... xj (n-1)) be the response characteristics when the code vector with the optimum gain is applied. It is represented by.

Xj = Hσyj (1) Here, the transfer function H (impulse response determinant H) is expressed by the following matrix.

[0102]

[Equation 1] The first line shows the filter output at t ₀ , the second line shows the filter output at t ₁ ···, and “σyj.” Shows the input.

At this time, the index j of the VQ code vector is determined so that the squared error between the target vectors x (n) and xj is minimized. when the jth code vector is selected
If the square error of x (n) and xj is Dj, Dj = | x (n) −xj | ² = | x (n) −H · σyj. | Represented by ^2, this calculation is performed by the second error calculation part 61b and the minimum square error computation section 63.

Here, the error with respect to all code vectors is calculated, and the index of the codebook that minimizes the value (that is, the optimum index) is determined. The LPC analysis part 11 performs LPC analysis (linear prediction analysis) from the input speech signal passing through the above LPC analysis window, and outputs the generated auditory weighting filter coefficients to the first and second auditory weighting filters 62a and 62b. Then, the number of reflection coefficients corresponding to the order of the short-term prediction filter 42 is sent to the reflection coefficient quantization section 23 and the pitch period extraction section 31 (the order of the short-term prediction filter is
(Nth order, N reflection coefficients are generated).

The reflection coefficient quantization section 23 is provided with a reflection coefficient table 24 provided in advance for N applied reflection coefficients.
Is quantized to the nearest reflection coefficient and sent to the short-term prediction filter coefficient generation unit 25. The optimum reflection coefficient index in the reflection coefficient table 24 at the time of quantization is transmission information.

Therefore, the short-term prediction filter coefficient generation part 25
Is used in the encoding process by converting the quantized reflection coefficient into a short-term prediction filter coefficient. On the other hand, the pitch period extraction unit 31 uses the input speech signal stored in the vector buffer 12 and the analysis result from the LPC analysis unit 11 to extract the basic pitch period of the input speech signal and generate the LPC prediction residual signal. And sends it to the long-term prediction coefficient quantizer 35.

The long-term prediction coefficient quantizer 35 calculates the long-term prediction filter coefficient using the basic pitch period and the LPC prediction residual signal, and then stores the calculated long-term prediction filter coefficient in the long-term prediction filter table 34 in advance. The quantized value is quantized to the closest quantized value.

The long-term prediction filter is represented by a third-order transfer function, and there are three filter coefficients. In the present invention, 3
Each filter coefficient is set as one vector, and the long-term prediction filter coefficient table 34 is set so that the error with this vector is minimized.
To explore. Further, the quantized long-term prediction filter coefficient is used in the encoding process.

Here, the voice coded information (VQ codebook index, reflection coefficient index, long-term prediction filter coefficient index) generated by these processes is subjected to format conversion for transmission to the line side by the multiplexing unit.

In FIG. 7, the detailed description of the processing procedure shown in FIG. 7 is omitted because it overlaps with the operation description of FIG. 6 described above, and in which part of FIG. 6 the processing of each step is performed will be described.

In the processing procedure shown in FIG. 7, one frame is composed of the first vector and the second vector.
Steps (S1) to 8 (S8) are the processing for the first and second vectors, and step 10 (S10) is the preparation for the next processing when the processing up to the second vector is completed (LPC). Analysis).

Therefore, the process does not proceed to step 10 until the processing for the first vector and the second vector is completed. Now, in step 1, the impulse response vector part 43
Initializes by performing impulse response calculation using the input long-term prediction filter coefficient, short-term prediction filter coefficient, auditory weighting filter coefficient, and pitch period.

In step 2, the vector buffer portion 12
Performs vectorization processing on the input speech signal to generate an m-sample speech vector signal s (n). Step 3
Then, the perceptual weighting filter 62a performs perceptual weighting on the voice vector signal obtained in step 2.

In step 4, 0 input response calculation (ZIR calculation) is performed for the long-term prediction filter 41, the short-term prediction filter 42, and the auditory weighting filter 62b connected in cascade. That is, 0, 0, 0, ... Is applied to the long-term prediction filter, the filter characteristic (response state) of each filter is calculated, and the calculation result is stored in the memory built in each filter.

In step 5, the auditory weighting filter 62
The output V (n) of a and the output r (n) of the auditory weighting filter 62b
The target vector x (n) is calculated by calculating the difference of. In step 6, the code vector extracted from the VQ codebook 51 is amplified by the amplification section 52 and the impulse response filter section.
An error between the impulse response vector xj output via 44 and the above target vector x (n) is calculated to obtain a comparison vector.

In step 7, the loop of VQ codebook 51 → impulse response filter 44 → error calculation part 61b → least square error calculation part 63 → VQ codebook 51 is used to index the VQ codebook that gives the least square error. Search for.

In step 8, if the optimum index of the VQ codebook can be searched by the encoding process, the code vector of the optimum index is passed through the gain portion 52 to the long-term prediction filter 41, short-term prediction filter 42, and perceptual weighting filter 62b. Apply to cause the response of these filters to be calculated. As a result, the calculation result of "0 state response" is obtained.

When performing the calculation in step 8, the calculation result obtained in step 4 is temporarily saved and the memory contents are cleared to "0". Then, the calculation result of "0 state response" and the saved memory content are added to update the memory content of each filter, but this update must be performed every time.

In step 9, if the second vector is not completed, the process returns to step 2, and if completed, the process proceeds to step 10. In step 10, a. “LPC analysis of the input data of the previous frame” and “adaptation of auditory weighting filter coefficients” are processed in the LPC analysis part 11, and b. “Adaptation of short-term prediction filter coefficients” is performed in the reflection coefficient quantization part. 23, reflection coefficient table 24, short-term prediction filter coefficient generation part 25, c. "Pitch cycle extraction" is processed by pitch cycle extraction part 31, d. "Long-term prediction filter coefficient adaptation" is long-term prediction coefficient quantum Processing section 35, long-term prediction filter coefficient table 34 ", e." Adjustment of gain "is processed by gain adaptor 53, and f." Impulse response calculation "is processed by impulse response vector section 43.

In FIG. 8, the received speech coded information is separated into a VQ codebook 51 index, a reflection coefficient table 26 index, a long-term prediction filter coefficient table 36 index and a pitch period in a separation unit 71.

Using the respective separated indexes, the corresponding code vector from the VQ codebook 51, the corresponding reflection coefficient from the reflection coefficient table 26, and the corresponding long-term prediction filter coefficient from the long-term prediction filter coefficient table 36 are respectively obtained. .

Further, the short-term prediction filter coefficient generator 25
A short-term prediction filter coefficient is generated using the input reflection coefficient. Then, the short-term prediction filter coefficient is given to the short-term prediction filter 42, and the long-term prediction filter coefficient is given to the long-term prediction filter 41 to which the pitch period is added.

On the other hand, the above code vector has a gain portion.
After the optimum gain indicated by the gain adaptor 53 is given at 52, decoded speech is generated and taken out through the long-term prediction filter 41 and the short-term prediction filter 42.

In FIG. 9, the received speech coded information is separated into a VQ codebook 51 index, a long-term prediction filter coefficient table 36 index and a pitch period in a separation unit 71.

Using each of the separated indexes, the corresponding code vector from the VQ codebook 51 and the corresponding long-term prediction filter coefficient are obtained from the long-term prediction filter coefficient table 36, respectively.

The filter coefficient of the short-term prediction filter 42 is obtained by subjecting the decoded speech signal output in the past by this filter to LPC analysis in the adaptor 46. Gain adaptor
53 predicts the gain to be given to the current code vector from the VQ code vector output from VQ code book 51 in the past and the gain added to it.

Therefore, the code vector taken out from the VQ codebook 51 is passed through the long-term prediction filter and the short-term prediction filter to which the pitch period and the long-term prediction filter coefficient are added after the gain part 52 is given an optimum gain. And outputs it as a decoded audio signal.

That is, according to the present invention, since the frame length and the vector length are short, the delay time due to encoding can be shortened. As a result, the buffer memory for storing various data necessary for the encoding process becomes small, the hardware scale can be reduced, and it contributes to the cost reduction and the power consumption reduction of the entire speech encoding apparatus. .

Further, by performing LPC analysis on the synthesized speech signal which is the output of the short-term prediction filter to generate the short-term prediction filter coefficient, the encoder does not need to transmit the information of the short-term prediction filter coefficient to the decoder side. , There is a possibility of lowering the bit rate.

Furthermore, in this method, the sound quality of the processed voice differs depending on the number of code vectors in the VQ codebook. Therefore, instead of reducing the information of the short-term prediction filter coefficient, it is possible to improve the sound quality at the same bit rate by increasing the number of code vectors of the VQ codebook.

Next, the description of FIGS. 10 to 20 will be made.
The basic processing procedure of FIG. 13 is equivalent to that of FIGS. 6 to 8 and has been described in detail above, so a brief description will be given. The major change is that an adaptive codebook is provided instead of the long-term filter in order to reproduce the basic pitch component, as will be described later.

Now, the operation of FIG. 10 will be briefly described with reference to the processing procedure of FIG. First, "0" is applied to initialize the program, and then the impulse response vector generation unit 116 performs impulse response calculation.

Here, the impulse response vector generator 11
In Fig. 6, a synthesis filter and a perceptual weighting filter are connected in cascade, and a synthesis filter coefficient and a perceptual weighting filter coefficient are given to these filters. Impulse is applied to this filter to obtain its output h (n). .

At this time, the impulse response filter 117 performs the convolution operation of the impulse response vector h (n) and the code vector to obtain the 0-state response vector y (n) (S1,
(See S2).

In addition, vectorization of the input audio signal in the vector buffer 101 (see S3), the perceptual weighting filter 10
Auditory weighted filtering of input vectors in 7b
(Refer to S4), Zero input response vector r (n) calculation of auditory weighted synthesis filter (106, 107a) and auditory weighting filter
Subtraction with the target vector V (n) which is the output of 107b (S5
Calculation of target vector x (n) in subtraction unit 102a
(See S6), pitch p search processing in adaptive codebook 109, determination of adaptive codebook index, adaptive codebook gain index from optimal gain β calculation,
Search process for code c in fixed codebook 112, determination of fixed codebook index and fixed codebook gain index from optimal gain γ calculation, update process of optimal codebook 109 (see S7 to S9), filter memory for each part Update with 106, 107a, 116, 117 (see S10).
Then, the program determines whether the frame processing is performed (S
(See 11).

This is because the linear prediction analysis 103, the reflection coefficient quantization unit 104a, and the filter coefficient generation unit 104b process each frame, but the other parts are processed in a vector manner, and the program determines this. Is.

For program processing, the linear prediction analysis unit
The linear prediction analysis of the input speech signal in 103 and the reflection coefficient quantization in the reflection coefficient quantization unit 12 are performed, and the reflection coefficient table 10
5, the reflection coefficient table index determination, the filter coefficient generation in the filter coefficient generation unit 13, the filter coefficient generation in the auditory weighting filters 107a and 107b, and the h (n) generation in the impulse response vector generation unit 116 are performed. Repeat (see S11-S15).

The operation of the decoder shown in FIG. 12 will be briefly described with reference to the processing procedure of FIG. In FIG. 13, similar to the encoder, the program is initialized and the transmission code is received (S
1, S2). After that, the optimum adaptive code vector and the optimum adaptive code vector gain corresponding to the received adaptive code book index and adaptive code book gain index are extracted from the adaptive code book 109 and the adaptive code book gain table 110, and the latter is obtained. It is given to the gain means 111 (see S3).

Further, the optimum fixed code vector corresponding to the received fixed codebook index and fixed codebook gain index and the optimum fixed code vector gain are set to fixed codebook 112 and fixed codebook.
From the gain table 113, the latter is gain means 11
Assign to 4 (see S4).

Then, the adaptive codebook 109 is updated and the memory in the synthesis filter 106 is updated to extract the output audio signal (see S5 and S6).
The reflection coefficient quantization unit 104a extracts the reflection coefficient corresponding to the received reflection coefficient table index from the reflection coefficient table and sends it to the filter coefficient generation unit 104b (see S7 and S8).

The filter coefficient generator 104b repeats the generation of the synthetic filter coefficient using the input reflection coefficient (see S9). The generated synthesis filter coefficient is sent to the synthesis filter 106.

By the way, it is premised that the audio signal is repeated in a certain cycle when the audio is reproduced. Therefore, in the configuration of the encoder shown in FIG. 6, the pitch period extraction unit 31 gives the pitch period extracted by using the output of the vector buffer to the long-term prediction filter 41, and periodically adds the pitch component to the output vector of the gain unit 52. ing. However, the most difficult point of this method is that it is difficult to find an appropriate pitch period when extracting the pitch period.

On the other hand, the adaptive codebook stores information indicating what kind of signal has been transmitted in the past. If there is a proper one among the stored signals, if it is taken out again, Pitch components can be combined.

Therefore, in the encoder shown in FIG. 10, an adaptive codebook 109 is provided instead of the long-term filter. That is, the adaptive codebook 109 is one in which the optimum codevectors generated in the past are accumulated in a buffer memory of a certain size (pitch extraction range + 1 vector length), and one vector is cut out from the accumulated data in the past. After the gain adjustment is performed with the optimum gain, a signal in the weighting region, that is, a weighting signal r (n) is generated by passing through the synthesis filter and the auditory weighting filter, and the difference from the input vector V (n) is evaluated.

Then, the next cut-out is shifted by one sample, and the same processing as above is repeated to find the one having the smallest error from the respective difference signals, and the position of the vector becomes the pitch, and as the optimum code vector, The index is transmitted as code information.

Adaptive codebook 109 uses the optimum codevectors generated in the past as adaptive codebook y0 in FIG.
As shown in (1), it is stored horizontally, and one vector is cut out from the position b, and the next cutting is performed by cutting out one vector from the position a shifted by one sample.

Here, the algorithm realized by the functional blocks in FIG. 10 is as follows. Synthesis filter 10
The transfer function F (z) of 6 is as follows.

[0148]

[Equation 2] The transfer function of the perceptual weighting filter 107 is as follows.

[0149]

(Equation 3) The evaluation formula D _P of the adaptive codebook search is as follows.

[0150] _{D p = MIN | vn -βHp |} 2 H:
Impulse response matrix The evaluation formula D _{c for} fixed codebook search is as follows. vn = vn − β _opt Hp _opt as D _c = MIN │ vn − γ Hc │ ² H: Impulse response matrix where z in the transfer function of the auditory weighting filter is the delay element of the filter, γ is a parameter, adaptive codebook search The vn − βHp part of the evaluation formula of is the part that finds the difference between the input signal (target vector) and the optimum value, and the evaluation formula of the fixed codebook search is the MIN of the adaptive codebook.
Since vn is a fixed value and this is used to calculate one's own vn MIN, it is a two-step process (two expressions).

Further, the impulse response determinant H is as follows.

[0152]

Hn = {h ₀ , h ₁ , h ₂ , ..., h _n-1 }: Impulse response sequence Note that this impulse response determinant and the above transfer function are the same equations, but Therefore, it is an impulse response matrix.

Next, FIGS. 14 to 18 will be described. In general, when storing data arranged in time series for array elements, the data is stored in order from the oldest time or the smallest subscript of the array. This is because it easily corresponds to the subscript expressed by the theoretical formula of the algorithm.

For example, when the vectorization of the input speech signal of FIG. 10 is expressed as follows, the vector s (n) = | s ₀ , s ₁ , s _2, ..., s _n-1 | Are sequentially stored from s ₀ to s _n−1 in the order of arrival (see FIG. 14 (a)).

In the present invention, as shown in FIG. 14 (b), the order on the time axis and the order of array subscripts are reversed. For that reason,
It is difficult to attach a correspondence to the theoretical formula, but it will be useful to reduce the amount of calculation at the time of implementation on an actual processor (
(See below).

Further, the impulse response filter 117 of FIG.
When synthesizing the 0-state response vector y (n) from the code vector, the convolution operation of the impulse response matrix H and the code vector y0 extracted from the adaptive codebook 109 is performed as shown in FIG.

As described above, the adaptive codebook 109 cuts out the data in the array one sample at a time and one vector at a time while shifting it. A convolution operation is sequentially performed on this vector, but in order to reduce the operation processing, a method of calculating only the difference at the time of the previous vector calculation is adopted.

Here, in the case where the oldest time is stored in the ascending order of the array subscripts and the order of the array elements is the same as the conventional time series, the convolution operation is as shown in FIG. Here, as an example, 1 vector = 10 samples,
When calculating in order from i = 0 to (n-1), for example, i =
When calculating (n-10), the part within the dotted line of i = (n-11) can be reused, so the leftmost column is subtracted for each row (difference calculation), i = n-
Calculate only the bottom 10 rows (difference calculation).

Similarly, when calculating i = (n-9), i = (n-10)
Since the part within the dotted line of can be diverted, the leftmost column is subtracted for each row and only the bottom row of i = n-9 is calculated. At this time, the number of calculations is the leftmost column: total: 9 times, subtraction: 9 times → difference calculation Bottom row calculation: total: 10 times, addition: 10 times → total difference calculation 38 times All four arithmetic operations are required. Also, at that time, the optimum codebook index to be selected is (n-
It is difficult to intuitively understand because i) needs to be calculated.

On the other hand, the storage order is reversed as shown in FIG. Further, in this case, the array element y0 0 is not used, and the case of performing the same position calculation is considered. i = 11
When a part of i = 10 is used in the calculation of, it is necessary to add only the leftmost column as shown in FIG.

The number of calculations at this time is 19 in total, that is, calculation in the leftmost column: integration: 10 times, addition: 9 times. In addition, the optimum adaptive codebook index at this time also has the current position n =
Since it is 0, the value of the cutout position i of the vector can be used as it is, which is intuitively easy to understand.

It should be noted that in FIG. 16, the calculation may be sequentially performed in the direction from a large value to a small value (backward to frontward). Calculation is performed in the order of (3) → (2) → (1) in Fig. 16.
For example, the data within the dotted line in (2) can be diverted to the process of (1), and the same process as in Fig. 18 can be performed, but the calculation of the optimal codebook index (difference calculation) can be performed using (ni) However, there remains a problem that the same processing as the conventional example is required. In addition to this, in order to reduce the calculation amount of the adaptive codebook, the property that there is a strong correlation between consecutive pitch information is used, and instead of searching the entire pitch extraction range each time, the pitch position is predicted in advance. , There is a method of performing a search process only for several samples before and after that.

As a generally used method, a method is used in which the pitch is calculated from the correlation calculation of the linear prediction residual waveform of the input signal (same as the long-term prediction filter) and used as the pitch candidate of the adaptive codebook.

In this case, for the preliminary search of the pitch, an inverse filter for obtaining the residual waveform from the input signal, a correlation calculating means for calculating the correlation, etc. are additionally required, which complicates the configuration. In the present invention, in the adaptive codebook search process, pitch candidates are set, and by sequentially updating,
No additional blocks occur. Figure about this method
It will be described with reference to FIGS.

FIG. 19 shows a method in which pitch extraction evaluation is performed for all vectors every time, and the adaptive codebook search is performed for each vector. Here, the number of vectors in one frame is m
And add 1 to m as each vector number. For each vector, the data in the adaptive codebook at that time is used to search the entire pitch extraction range to find the optimum pitch from P1 to Pm. In this case, the processing amount of the search operation in each vector is the same because the extraction range is the whole (see FIGS. 19 (a) to 19 (c)).

Here, it is generally known that the pitch period of a voice has a strong correlation between before and after successive voices and takes a substantially constant value. In other words, P1 and P2, P2 and P3 ... Therefore, a method of limiting the search range and reducing the amount of calculation by utilizing this property will be described.

In FIG. 20, for the first vector,
Similar to the above, the entire pitch range is searched to find P1. In the next second vector, assuming that it should be near P1, the search range is limited to the range of P1 ± k, and P2 is obtained.

Note that, for the selection of K, for example, a method of setting the average or maximum value of the difference between before and after from P1 and Pm obtained in the explanation of FIG. 17 from the result of evaluation of a plurality of voice samples can be considered. .

Similarly, for the third and subsequent vectors, the amount of calculation can be reduced by performing the search process within the range of the pitch ± K of the previous vector (see FIGS. 20 (a) to 20 (c)). Further, the process of searching the entire pitch range in a frame is not limited to once for the first vector, but once for two vectors and once for three vectors.
It is also possible to consider a method of performing the full pitch search once or multiple times. In this case, the number of calculations increases, but the accuracy of pitch search increases, which is effective in improving the characteristics.

The above is a method of leaving one pitch candidate, but a method of leaving a plurality of pitch candidates (two or more candidates) and performing a search process within a range of ± K for each candidate. Can also be considered. The search process for a certain candidate is the same as in FIG. 20, and for example, PI is P1-1, P1-2
・ You can think that there is more than one.

[0171]

As described above in detail, according to the present invention, since the delay time due to the encoding can be shortened, the buffer memory for storing various data necessary for the encoding processing becomes small and the hardware can be reduced. The wear scale can be reduced.

Further, by performing LPC analysis on the synthesized speech signal which is the output of the short-term prediction filter to generate the short-term prediction filter coefficient, the encoder does not need to transmit the information of the short-term prediction filter coefficient to the decoder side. Further, there is an effect that a lower bit rate or higher sound quality can be achieved and the performance of the high-efficiency audio transmission device can be improved.

Further, by adopting the coding method according to the present invention, it is possible to provide wideband speech at a transmission rate lower than that of the conventional example and to provide a more natural and realistic environment. In addition, by reversely processing the order of the convolutional operations, it is possible to reduce the amount of operations required for the processing and to reduce the scale of realized hardware.

[Brief description of drawings]

FIG. 1 is a functional explanatory diagram of a main part of a speech coder according to first to third aspects of the present invention.

FIG. 2 is a functional explanatory diagram of a main part of a speech coder according to a fourth aspect of the present invention.

3 is an explanatory diagram of a processing timing of FIG.

FIG. 4 is an explanatory diagram of an LPC analysis window

FIG. 5 is a functional explanatory diagram of a main part of a speech encoder / decoder according to seventh to twelfth aspects of the present invention.

FIG. 6 is a functional explanatory diagram (encoder) of the first to third embodiments of the present invention.

7 is an explanatory diagram of a processing procedure of FIG.

8 is a functional explanatory diagram of a decoder corresponding to the encoder of FIG.

FIG. 9 is a functional explanatory diagram of a sixth embodiment of the present invention (decoder).
It is.

FIG. 10 is a functional explanatory diagram (encoder) of seventh to ninth and eleventh embodiments of the present invention.

11 is an explanatory diagram of a processing procedure of FIG.

FIG. 12 is a functional explanatory diagram (decoder) of the tenth and twelfth embodiments of the present invention.

13 is an explanatory diagram of a processing procedure of FIG.

FIG. 14 is a diagram for explaining a data storage order, in which (a) is a normal order storage, and (b) is a reverse order storage, which is the case of the seventh aspect of the present invention.

FIG. 15 is an explanatory diagram of the calculation amount in the case of FIG. 14 (1).
It is.

FIG. 16 is an explanatory diagram of the calculation amount in the case of FIG.
It is.

FIG. 17 is a calculation amount explanatory diagram (1) of the seventh embodiment of the present invention.

FIG. 18 is a diagram (No. 2) for explaining the amount of calculation in the seventh embodiment of the present invention.

FIG. 19 is an explanatory diagram of a pitch preliminary search method for an adaptive codebook, in which (a) is a search method illustration and (b) and (c) are processing procedure illustrations.

FIG. 20 is an explanatory diagram of an eighth embodiment of the present invention, (a) is a search method explanatory diagram, and (b) and (c) are processing procedure explanatory diagrams.

FIG. 21 is an example of a main part configuration diagram of an audio signal generation model.

FIG. 22 is an explanatory diagram of vectors / frames.

FIG. 23 is an explanatory diagram of processing timing of a conventional encoder.

[Explanation of symbols]

 11, 103 Linear prediction (LPC) analysis 12, 101 Vector buffer 21 Reflection coefficient codebook 22 Adaptor 23, 104a Reflection coefficient quantizer 24, 105 Reflection coefficient table 25 Short-term prediction filter coefficient generation 31 Pitch period extractor 32 Adaptor 33 Long-term prediction coefficient codebook 34 Long-term prediction filter coefficient table 35 Long-term prediction coefficient quantization 41 Long-term prediction filter 42 Short-term prediction filter 43 Impulse response vector 44, 117 Impulse response filter 51 VQ codebook 52, 111, 114 Gain unit 53 Gain adaptation 61 Error calculation part 62, 107 Auditory weighting filter 63, 108 Minimum error square calculation 109 Adaptive codebook 112 Fixed codebook 116 Impulse response vector generator 106 Synthesis filter 110 Adaptive codebook gain table 113 Fixed codebook gain table

Front page continuation (72) Inventor Shuei Eguchi 3-22-8 Hakataekimae, Hakata-ku, Fukuoka City, Fukuoka Prefecture Fujitsu Kyushu Digital Technology Co., Ltd. (72) Inventor Kyoji Ota Kami, Nakahara-ku, Kawasaki City, Kanagawa Prefecture 4-1-1 Odanaka, Fujitsu Limited (72) Inventor Masanao Suzuki 4-1-1, Kamiodanaka, Nakahara-ku, Kawasaki-shi, Kanagawa Inside Fujitsu Limited

Claims (13)

[Claims] 1. Before performing high efficiency coding on an input speech signal of an Nth frame (N is a positive integer), (N-
1) Using an audio information parameter generated by performing a linear prediction analysis on an input audio signal of a past frame including the n-th frame, an encoding process is performed on the input audio signal of the N-th frame. A highly efficient voice transmission method characterized by the above. 2. The linear prediction analysis window for weighting an input speech signal as much as a frame closer to the N-th frame among the past frames is used. Efficient voice transmission method. 3. A voice information generating means for generating a voice information parameter by performing a linear prediction analysis of an input voice signal,
Multiple noise signal patterns that are sound sources are stored,
Vector quantization codebook means for giving a gain to the extracted noise signal pattern and outputting the noise signal pattern, and a long-term prediction filter having the applied voice information parameter as a filter coefficient,
A short-term filter, a speech synthesis filter means for generating a synthesized speech signal by using the output of the vector quantization codebook means, and an error calculation between the input speech signal and the generated synthesized speech signal are repeated to obtain a sound source with a minimum error. In a high-efficiency speech coder having a minimum error square calculation means for determining a number, speech information generated by performing linear prediction analysis on an input speech signal of a past frame including the (N-1) th frame A high-efficiency speech coder, which is configured to perform an encoding process on an input speech signal of an N-th frame using parameters. 4. The voice information parameter generated by performing a linear prediction analysis on a synthetic speech signal output in the past is used as a filter coefficient of the short-term prediction filter. High efficiency voice encoder. 5. A vector quantizer which separates input speech coded information to extract a sound source number and a sound information parameter, and a noise signal pattern corresponding to an applied sound source number and outputs the noise signal pattern with a gain. In a high-efficiency speech decoder including a codebook means, a long-term prediction filter having a filter characteristic corresponding to an applied speech information parameter, and a short-term prediction filter, the short-term prediction filter converts a decoded speech signal output in the past. A high-efficiency speech decoder characterized in that the speech information parameter generated by performing linear prediction analysis is used as a filter coefficient. 6. The high-efficiency speech coder according to claim 4,
A high-efficiency voice transmission device, characterized in that it encodes and transmits an input voice signal, and receives and decodes a voice signal encoded by using the high-efficiency voice decoder according to claim 5. 7. When performing a convolution operation of an impulse response matrix and a code vector with a synthesis filter, the operation sequence is started from the most recent code vector, and the operation is sequentially shifted to the past code vector to perform the convolution operation. An arithmetic method characterized by performing. 8. When the pitch search of the adaptive codebook is performed in units of m (m is a positive integer) vector that constitutes a frame, the first vector is searched over the entire range to obtain the optimum pitch P _1. , The second vector to the m-th vector are characterized in that the optimum pitch is found by limiting the pitch search within a preset range centered on the optimum pitch found in the previous pitch search. Calculation method. 9. A zero state is obtained by performing a convolution operation of the added code vector obtained by adding the code vectors respectively taken out from the adaptive codebook means and the fixed codebook means and the impulse response vector from the impulse response vector generation means. A high-efficiency speech code which is provided with impulse response filter means for obtaining a response vector, and searches for a combination of optimum code vectors such that the error between the 0-state response vector from the impulse response filter means and the input target vector is minimized. In the coding method, a high-efficiency speech coding method is characterized in that a convolution calculation is performed using the calculation method of claim 7 and a pitch search is performed using claim 8. 10. A high-efficiency speech decoding method, which decodes a signal encoded by the speech encoding method according to claim 9 to reproduce an original speech signal. 11. A high-efficiency speech coder characterized in that a convolution operation is performed by using the operation method according to claim 7, and a pitch search is performed by using claim 8. 12. A high-efficiency speech decoder having a configuration for decoding a signal encoded by the speech encoding method according to claim 9 to reproduce an original speech signal. 13. An input speech signal is encoded and transmitted using the high efficiency speech coder of claim 11, and a speech signal encoded by the high efficiency decoder of claim 12 is received and decoded. A high-efficiency voice transmission device having the above structure.