JPH08211895A - System and method for evaluation of pitch lag as well as apparatus and method for coding of sound - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a compact and accurate pitch evaluating system into which a multiple resolution analysis to encode a voice is incorporated. SOLUTION: The pitch evaluating device and method evaluate pitch lag of a voice by using a multiple resolution system. This system contains a step of sampling a voice, a step of alternately applying discrete Fourier transform and a step of squaring a result. Next, DTF is performed to convert a voice sample into a separate area to squared amplitude. Next, an initial pitch lag is obtained with low resolution. After the pitch lag is evaluated with low resolution, algorithm made precise is on the basis of minimizing an anticipating error in a time area. Next, a pitch lag made preceise can be directly used in encoding a voice.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION Signal modeling and parameter estimation play an increasingly important role in data compression, decompression, and coding. To model the basic speech sound, the speech signal must be sampled as a discrete waveform and processed digitally. In one type of signal coding technique called linear predictive coding (LPC), the signal value at some particular time index is modeled as a linear function of the previous value. Thus, the latter signal is linearly predicted according to the previous values. as a result,
The effective signal representation can be determined by evaluating and applying certain prediction parameters to represent the signal. Currently, LPC techniques are used for speech coding including Code Excited Linear Prediction (CELP).

[0002] Pitch information is accepted as a solid sound indicator and indication for coding purposes. Pitch describes the basic features or parameters of the speaker's voice. Since human speech is generally not easily mathematically quantifiable, speech evaluation models that can effectively evaluate speech pitch data provide more accurate and accurate encoded and decoded speech. However, some CELP
(Eg vector sum excitation linear prediction (VSELP),
In current speech coding models, such as multi-pulse, regular pulse, algebraic CELP, etc.) and MBE coder / decoders (“codecs”), pitch estimation algorithms require high accuracy and low complexity. Therefore, pitch evaluation is often difficult.

Several pitch lag evaluation schemes are used in connection with the codecs (time domain, frequency domain, and cepstrum domain) described above. Due to the close relationship between pitch lag and voice reproduction, the accuracy of pitch estimation has a direct impact on speech quality. CEL
In P-coder, speech generation is based on predictions (long-term pitch prediction and short-term linear prediction). FIG. 1 shows a block diagram of voice reproduction by a typical CELP coder.

In order to compress audio data, it is desirable to extract only essential information and avoid redundant transmission. Speech can be classified into short blocks, where typical parameters can be identified in every block. FIG.
As shown in, the CELP speech coder generates an LPC parameter 110 and a pitch lag parameter 112 (including the lag and its coefficient) in order to generate a good quality speech.
An optimal new code vector 114 with its gain parameter 116 must be extracted from the input speech to be encoded. The coder quantizes the LPC parameters by implementing the appropriate coding scheme. The quantization index for each parameter includes the information to be stored or transmitted to the speech decoder. In the CELP codec, pitch prediction parameters (pitch lag and pitch coefficient) are determined in the time domain, but in the MBE codec,
The pitch parameter is evaluated in the frequency domain.

After the LPC analysis, the CELP encoder
Determine the appropriate LPC filter 110 for the current speech coded frame (typically taken at about 10-40 ms). The LPC filter is represented by the following equation.

[0006]

[Equation 1]

In this equation, np is the LPC prediction order (usually about 10), y (n) is the sampled voice data, and n is the time index. L above
The PC equation describes the evaluation of the current sample according to a linear combination of past samples. A hearing correction filter based on the LPC filter modeled on the sensitivity of the human ear is defined here by:

[0008]

[Equation 2]

To extract the desired pitch parameter, the pitch parameter that minimizes the next weighted coding error energy must be calculated for each coded subframe, where one coded frame is It can be divided into several coding subframes for analysis and coding.

[0010]

(Equation 3)

In this equation, T is the target signal representing the perceptually filtered input signal and H is the filter W.
It is an impulse response matrix of (z) / A (z). P _Lag
Is the pitch prediction contribution with the pitch lag “Lag” and the prediction coefficient β uniquely defined for a given lag, and C _i is the codebook contribution related to the index i in the codebook and its corresponding gain α. . Typically, the pitch of human speech varies between 2ms and 20ms. Therefore, when speech is sampled at a sampling rate of 8 KHz, the pitch lag roughly corresponds to 20 to 147 samples. Furthermore, i is 0
And Nc−1, where Nc is the size of the new codebook.

Consider a one-tap pitch predictor and one new codebook. However, the typical shape of the pitch predictor is typically a multi-tap scheme, and the typical shape of the new codebook is multi-level vector quantization, or multiple new codebooks are used. More specifically, in speech coding, the one-tap pitch predictor indicates that the current speech sample can be predicted by one past speech sample, while the multi-tap predictor indicates that the current speech sample has more than one past speech sample. It means that it can be predicted by the sample.

Due to concerns about complexity, schemes that are suboptimal have been used in speech coding schemes. For example, pitch lag evaluation is 2.5ms to 1
It may be done by simply evaluating the possible lag values in the range between the L ₁ and L ₂ samples to cover 8.5 ms. Therefore, the estimated pitch lag value is determined by maximizing:

[0014]

[Equation 4]

Although this time domain method can determine the true pitch lag, for female voices with high pitch frequencies, the pitch lag determined by equation (1) can be a multiple of the true lag rather than the true lag. Correct the evaluation error (for example, smooth lag) to avoid this evaluation error.
An additional process is required, which in turn causes undesired complexity.

However, such overcomplexity is a significant drawback when using the time domain method.
For example, in order to determine the lag using only an integer number of lags, the time domain method is 3 million operations per second (3 MO
At least P) is required. Moreover, if pitch lag smoothing and fractional pitch lag are used, the complexity would be approximately 4 MOPs. In practice, approximately 6 million Digital Signal Processing Machine Instructions (6 DSP MIPs) per second are required to perform a full range pitch lag evaluation with acceptable accuracy. Therefore, it is generally accepted that pitch evaluation requires 4 to 6 DSP MIPs. There are other schemes that can reduce the complexity of pitch estimation, but such schemes often sacrifice quality.

MBE, an important element in the class of sine coders
In the coder, the coding parameters are extracted and quantized in the frequency domain. The MBE voice model is shown in FIGS. In the MBE speech encoder / decoder ("Bogodha") described in FIGS. 2 and 3, the fundamental frequency (or pitch lag) 210, the voiced / unvoiced decision 21
2, and the spectral envelope 214 is extracted from the input speech in the frequency domain. The parameters are then quantized and encoded into a bitstream that can be stored or transmitted.

In MBE vocoders, the fundamental frequency must be evaluated with high accuracy in order to achieve good quality speech. The fundamental frequency is evaluated in two stages. First
, The initial pitch lag is searched within the range of 21 to 114 samples, and the input speech 21
6 and the synthesized speech 218 by minimizing the weighted mean square error equation 310 (FIG. 3).
2.6 ms to 14.000 at a sampling rate of 000 Hz.
Covers 25 ms. The mean squared error between the original and synthesized speech is given by:

[0019]

(Equation 5)

In this equation, S (ω) is the original speech spectrum, S ^ (ω) (^ is considered to be above the capital S) is the synthesized speech spectrum, and G
(Ω) is a frequency-dependent weighted function. As shown in FIG. 4, the pitch tracking algorithm 410 is used to update the initial pitch lag estimate 412 by using the pitch information of adjacent frames.

This method is used because of the assumption that the fundamental frequency should not change abruptly between adjacent frames. Pitch estimates of two past adjacent frames and two future adjacent frames are used for pitch tracking. The mean squared error (which includes two past frames and two future frames) is then minimized to find a new pitch lag value for the current frame. After tracking the initial pitch lag, pitch lag multiple inspection scheme 414 is applied to remove multiple pitch lags to smooth the pitch lags.

Referring to FIG. 4, the second of the fundamental frequency evaluations
Pitch lag refinement 416 is used at the stage to increase the accuracy of pitch evaluation. Pitch lag candidate values are formed based on the initial pitch lag estimate (ie, new pitch lag candidate values are formed by adding or subtracting a fraction from the initial pitch lag estimate). Therefore, a refined pitch lag estimate 418 can be determined among the pitch lag candidates by minimizing the mean square error function.

However, frequency domain pitch estimation has certain drawbacks. First, the complexity is very high. Second,
The pitch lag must be searched within 20 and 114 samples covering only 2.5 ms to 14.25 ms, limiting the window size to 256 samples to accommodate a 256-point FFT. However, it is not possible to collect a sufficient number of samples within the 256 sample window for very low pitch frequency speakers or for speech with pitch lags greater than 14.25 ms. Moreover, only the averaged pitch lag is evaluated over the speech frame.

In 1967, A. M. Nor (AMNol
Another modified method was proposed using the cepstrum region pitch lag estimation (Fig. 5) proposed by l).
The cepstral range pitch lag evaluation samples approximately 37 ms of speech at 510, thus covering at least two periods of the maximum possible pitch lag (eg, 18.5 ms). The 512-point FFT is then applied (at block 512) to the speech frame window to obtain the frequency spectrum. Taking the logarithmic 514 amplitude of the frequency spectrum, another 512 point inverse FF
T516 is applied to obtain the cepstrum. The weighting function 518 is applied to the cepstrum and the cepstrum peaks are detected at 520 to determine the pitch lag. Next, the tracking algorithm 522 is executed,
Remove any pitch multiples.

However, the cepstral pitch detection method has some drawbacks. For example, there are high computational demands. To cover the range of pitches between 20 and 147 samples at a sampling rate of 8 KHz, the 512 point FFT has to be performed twice. The cepstrum pitch estimate provides only an estimate of the averaged pitch lag over the analysis frame, so the accuracy of the estimate is insufficient. However, for low bit rate speech coding, it is important that the pitch lag value be evaluated over a short period of time. As a result, cepstral pitch estimation is rarely used today for high quality, low bit rate speech coding. Therefore, due to the limitations of each of the schemes described above, it is desirable for the means for effective pitch lag estimation to meet the need for high quality, low bit rate speech coding.

[0026]

SUMMARY OF THE INVENTION Accordingly, it is an object of the present invention to provide a pitch estimation system incorporating a multi-resolution analysis for speech coding which requires minimal complexity and high accuracy. . In a particular embodiment, the present invention is directed to a speech coding apparatus and method using CELP technology and various other speech coding and recognition systems. Thus, better results are provided with less computational means, while maintaining the required high accuracy.

These and other objects are achieved, in accordance with an embodiment of the present invention, by a pitch lag evaluation scheme which enables accurate reproduction and reproduction of speech quickly and effectively. The pitch lag is extracted for a given speech frame and then refined for each subframe.
After the minimum number of speech samples has been obtained by directly sampling the speech, the discrete Fourier transform (DF
T) is applied and the resulting amplitude is squared.
The second DFT is then performed. Therefore, the exact initial pitch lag for speech samples in a frame is
It is possible to determine between a possible minimum of 20 samples and a maximum lag value of 147 samples at a sampling rate of 8 KHz. After obtaining the initial pitch lag estimate, time domain refinement must be performed for each subframe to further improve the accuracy of the estimate.

[0028]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT The pitch lag evaluation scheme according to the preferred embodiment of the present invention is generally shown in FIGS.
And 9 are shown. First, N speech samples {x (n), n = 0, 1, ..., N-1} are collected.
(Step 602 of FIG. 6) N is, for example, 8000 Hz
May be equal to 320 voice samples to accommodate a typical 40 ms voice window at a sampling rate of. The value of N is determined by a roughly estimated speech period, where at least two periods are generally needed to generate the speech spectrum. Thus, N
Must be greater than twice the maximum possible pitch lag, where {x (n), n = 0, 1, ..., N−
1}. Moreover, a Hamming window 604 or other window covering at least two pitch periods is preferably implemented.

[0029]

(Equation 6)

According to an embodiment of the invention, the function G (f)
It is recognized that the logarithm of G (f) differs from C (n) with the conventional cepstrum transform used in equation (4) instead. The reason for this difference is generally complexity.
It is desirable to reduce complexity by removing logarithmic functions that would require substantially more computational resources if not eliminated. Furthermore, cepstrum or C (n)
Comparing pitch lag evaluation methods using functions, different results were obtained only for unvoiced or transitional sections of speech. For example, the definition of pitch is unclear for unvoiced or transitional speech. It has been said that transition speech has no pitch, but it is also said that some prediction can always be shown to minimize error.

Thus, once C (n) is determined (step 610), the pitch lag for a given speech frame can be determined at step 614 by solving the following equation:

[0032]

(Equation 7)

In this equation, arg [•] determines the variable n that satisfies the internal optimization function, and L ₁ and L ₂ are defined as the smallest possible pitch lag and the largest possible pitch lag, respectively. For speech coding convenience, the difference between L ₂ and L ₁ is preferably a power of 2 for binary representation. In the preferred embodiment, L ₁ and L ₂ take values of 20 and 147, respectively, which are in the typical human voice pitch lag range of 2.5 ms to 18.375 m.
s, where the spacing between L ₁ and L ₂ is a power of 2. W (i) is a weighted function, 2M + 1
Represents the window size. Preferably, {W (i)
= 1, i = 0, 1, ..., 2M}, and M = 1.

Although the resulting pitch lag is an averaged value, it has been found to be reliable and accurate. The effect resulting from averaging is due to the absolutely large analysis window size, for a lag of 147 samples, the window size should be at least twice the lag value. Unfortunately, however, a signal from one speaker, such as a female speaker, which typically exhibits a small pitch lag, may include 4 to 10 pitch periods in such a large window. If there is a change in pitch lag, the proposed pitch lag estimate will only produce an average pitch lag. As a result, the use of such averaged pitch lags in speech coding can cause significant degradation in speech evaluation and playback.

Because of the relatively fast changes in pitch information in speech, most speech coding systems based on the CELP model evaluate and transmit pitch lag once every subframe. Thus, in CELP-type speech coding, where one speech frame is divided into several speech subframes, which are typically 2 ms to 10 ms long (16 to 80 samples), the pitch information is updated in each subframe. To be done. Therefore, accurate pitch lag values are needed only for subframes. However, the pitch lag evaluated according to the above scheme is not accurate enough for accurate speech coding due to the effects resulting from averaging.

Thus, in a particular embodiment of the invention, a refined search based on the initial pitch lag estimate is performed in the time domain to improve the accuracy of the estimate (step 618). A simple autocorrelation method is performed for a particular coding period or subframe with approximately averaged Lag values.

[0037]

(Equation 8)

In this equation, arg [•] determines the variable n that satisfies the internal optimization function, k is the first sample of the subframe, l is the refinement window size, and m is the search range. is there. In order to determine the exact pitch lag value, the refinement window size should be at least 1 pitch period. However, the window should not be too large to avoid the effects of averaging. For example, preferably 1 = Lag + 10, and m = 5. Thus, according to the time domain refinement of equation (6), a more accurate pitch lag can be evaluated and applied to subframe coding.

In operation, the fast Fourier transform (FF
Although T) may be more computationally efficient than a general DFT, the drawback with FFT is that the window size must be a power of two. For example, 1
It has been shown that the maximum pitch lag of 47 samples is not a power of 2. To include the maximum pitch lag,
A window size of 512 samples is required. However, this results in poor quality of the pitch lag estimate for female voices due to the effects resulting from the averaging described above and requires a large amount of computation. If a window size of 256 samples is used, the effects resulting from averaging are reduced and complexity is reduced. However, such a window cannot handle pitch lags greater than 128 samples in speech.

To overcome some of these problems, an alternative preferred embodiment of the present invention utilizes a 256-point FFT to reduce complexity and a modified signal is used to evaluate pitch lag. It is the downsampling process that modifies the signal. Referring to FIGS. 7 and 8, N speech samples {x (n), n = 0, 1,
, N-1} are collected (step 702), where N is greater than twice the maximum pitch lag. Next, the N speech samples are downsampled into 256 new analysis samples using linear interpolation according to the following equation (step 704).

[0041]

[Equation 9]

In this equation, λ = N / 256,
The value in square brackets, that is, [i.lambda.] Indicates the maximum integer value of i.lambda. Or less. Then, in step 705, a Hamming window or other window is applied to the interpolated data.

In step 706, the pitch lag evaluation is 2
Performed over y (i) using a 56-point FFT to generate the amplitude Y (f). Next, steps 708 to 710 are performed as described with respect to FIG. However, G (f) is further filtered (step 709) to reduce high frequency components of G (f), which is not useful for pitch detection. Once y
(I) If the lag i.e. Lag _y of Rarere calculated according to equation (5) (step 714), which determines the pitch lag estimation is rescaled in step 716.

[0044]

[Equation 10]

In summary, the above procedure for determining the initial pitch estimate for a coded frame is as follows.

(1) A standard 40 ms coded frame is subdivided into pitch subframes 802 and 804. Each pitch subframe is approximately 20 ms long.

(2) N = 3 so that the pitch analysis window 806 is positioned at the center of the last subframe.
Take 20 speech samples and find the lag for that subframe using the proposed algorithm.

(3) Determine the initial pitch lag value for the pitch subframe. Next, time domain refinement is performed at step 718 over the original speech sample x (n). Thus, in the embodiment of the present invention, the pitch lag value can be accurately evaluated while reducing the complexity and yet maintaining high accuracy. With the FFT embodiment of the present invention, it is not difficult to handle pitch lag values greater than 120.

More specifically, the time domain refinement is based on the original speech support.
It is done over the sample. For example, the 40ms mark
As shown in Fig. 9, the coding frame is composed of eight 5m frames.
It is divided into s subframes 808. Early pitcher
Evaluation lag₁And lag ₂Is the current encoding
The last encoded subframe of each pitch subframe in the
It is a lag evaluation for a rame. lag₀Is the preceding sign
Refinement of the second pitch sub-frame in an optimized frame
It is a lag evaluation. lag₁, Lag₂, And l
ag₀The relationship between is shown in FIG.

Initial pitch lags lag ₁ and lag ₂
Is first refined according to the following equation to improve its accuracy (step 718 of FIG. 8).

[0051]

[Equation 11]

Here, N _i is an index of the start sample in the pitch subframe with respect to the pitch lag _i .
Preferably M is selected to be 10 and L is lag _i +10
And i indicates the pitch subframe index.

Once the initial pitch lag refinement is complete
And the pitch lag of the encoded subframe can be determined. Mark
Pitch lag of encoded subframe is lag₁, Lag₂,
And lag₀Evaluated by linearly interpolating
Be done. Accuracy of pitch lag estimation for coded subframes
Is interpolated for each encoded subframe according to the following steps
It is improved by refining the pitch lag. precision
Initialized pitch evaluation lag₁, Lag₂,and
lag₀Interpolated pits for coding subframes based on
Chirag is {lag_I(I), i = 0, 1, ..., 7} is a table
If you do, lag _I(I) is determined by the following equation.

[0054]

(Equation 12)

Further refinement will be needed as the accuracy of the pitch lag estimation provided by linear interpolation is not sufficient. Given pitch lag rating {lag
_{For I} (i), i = 0, 1, ..., 7}, each lag _I
(I) is further refined by the following equation (step 7)
22).

[0056]

(Equation 13)

Here, Ni is an index of the starting sample in the coded subframe for the pitch lag (i). In the example, M is chosen to be 3 and L equals 40.

Furthermore, linear interpolation of pitch lag is important in the unvoiced section of speech. The pitch lag obtained by some analysis method tends to be arbitrarily distributed to unvoiced speech. However, due to the relatively large pitch sub-frame size, if the lag for each sub-frame is too close to the originally determined sub-frame lag (determined in step (2) above), then the audio will originally be The unwanted artificial periodicity that was not present is added. In addition, linear interpolation easily solves the problems associated with poor quality unvoiced speech. Furthermore, since the lag of subframes tends to be arbitrary, once interpolated, the lag for each subframe is also very arbitrarily distributed, which guarantees speech quality.

[Brief description of drawings]

FIG. 1 is a block diagram of a CELP voice model.

FIG. 2 is a block diagram of an MBE voice model.

FIG. 3 is a block diagram of an MBE encoder.

FIG. 4 is a block diagram of pitch lag evaluation in the MBE vocoder.

FIG. 5 is a block diagram of a pitch lag detection method based on cepstrum.

FIG. 6 is an operational flow diagram of pitch lag evaluation according to an embodiment of the present invention.

FIG. 7 is a flow diagram of pitch lag evaluation according to another embodiment of the present invention.

FIG. 8 is a flow diagram of pitch lag evaluation according to another embodiment of the present invention.

FIG. 9 is a diagram of speech encoding according to the embodiment of FIG.

[Explanation of symbols]

 802 pitch subframe 804 pitch subframe 806 pitch analysis window 808 subframe

Front Page Continuation (72) Inventor Tom Hong Lee, 501, Knowwood Drive, Middletown, NJ, 07748, USA

Claims (20)

[Claims] 1. A system for estimating pitch lag for speech quantization and compression, wherein speech is defined by a plurality of speech samples, wherein the estimation of the current speech sample follows a linear combination of past samples. And the system includes means for applying a first discrete Fourier transform (DFT) to the audio samples, the first DFT having an associated amplitude, and further comprising: Means for squaring the amplitude, means for applying a second DFT to the squared amplitude, means for determining an initial pitch lag value according to a time domain transformed speech sample, Means for encoding said speech samples according to a refined pitch lag value, and a system for evaluating pitch lag for speech quantization and compression. Temu. 2. The initial pitch lag value has an associated prediction error, and the system further includes means for refining the initial pitch lag value, the associated prediction error being minimized. The system according to Item 1. 3. Means for classifying the plurality of speech samples into a current coded frame, means for dividing the coded frame into a plurality of pitch subframes, and the pitch subframe with a plurality of codes. Means for subdividing into coded subframes, and initial pitch lag estimates lag ₁ and la representing lag estimates, respectively, for the last coded subframe of each pitch subframe in the current coded frame.
means for evaluating the g _2, and means for refining the pitch lag estimation lag ₀ of the second pitch subframe in the previous coding frame, lag _1, lag _2, and linearly interpolating lag ₀ The system of claim 1, further comprising means for evaluating a pitch lag value for the encoded subframes and means for further refining the interpolated pitch lag for each encoded subframe. 4. The system of claim 1, further comprising means for down-sampling the audio samples to a down-sampling value for a schematic representation with a small number of samples. 5. The initial pitch lag value is calculated by the equation (Lag
System according to claim 4, corrected by _scaled = number of audio samples / downsampled value). 6. The system of claim 1, wherein the means for refining the initial pitch lag value comprises autocorrelation. 7. Speech input means for receiving said speech samples, a computer for processing said refined pitch lag value and reproducing the input speech as encoded speech, said encoded speech. The system according to claim 1, further comprising: an audio output means for outputting the. 8. A speech coder for reproducing and encoding input speech, said speech coder inducing linear predictive coding (LPC) parameters and speech reproduction for generating speech. A new codebook representing a plurality of vectors referred to in the above is used, and the speech coding apparatus comprises a speech input means for receiving the input speech, and a computer for processing the input speech. Including and
The computer has means for cutting out a current coded frame in the input speech, means for dividing the coded frame into a plurality of pitch subframes, and a pitch analysis window having N speech samples. And a means for defining an initial pitch lag value for each pitch subframe, and the pitch analysis window extending for each pitch subframe. Means for dividing a frame into a plurality of coded subframes, the initial pitch lag estimate for each pitch subframe being a lag for the last coded subframe of each pitch subframe in the current coded frame. Representing an evaluation, the computer provides the evaluated pitch lag value to the Means for linearly interpolating between coded subframes and determining a pitch lag estimate for each coded subframe, and means for refining the linearly interpolated lag value of each coded subframe And a speech coding apparatus for reproducing and coding an input speech, the apparatus further comprising speech output means for outputting speech reproduced according to the refined pitch lag value. 9. The computer uses a downsampled value X to represent a small number of samples.
Further comprising means for down-sampling the N speech samples, and means for correcting the pitch lag value such that the corrected lag value Lag _scaled = N / X.
The device according to claim 8. 10. The apparatus of claim 8, further comprising sampling means for sampling the input speech at a sampling rate R, the N speech samples being determined according to the equation N = R * X. 11. The apparatus according to claim 10, wherein X = 25 ms, R = 8000 Hz, and N = 320 samples. 12. Each coded frame is approximately 40 ms
9. The device of claim 8 having a length of. 13. A method for estimating pitch lag for speech quantization and compression, wherein said speech is defined by a plurality of speech samples, wherein the estimation of the current speech sample follows a linear combination of past samples. Determined in the domain, the method comprising: applying a first Discrete Fourier Transform (DFT) to the audio samples, the first DFT having an associated amplitude, and further comprising the amplitude of the first DFT. Squared, applying a second DFT to the squared amplitude of the first DFT, and determining an initial pitch lag value according to a time domain transformed speech sample, The initial pitch lag value has an associated prediction error, and the method further comprises: refining the initial pitch lag value using autocorrelation. Look, the associated prediction error is minimized,
Further, a method for estimating pitch lag for speech quantization and compression, comprising encoding the speech samples according to the refined pitch lag value. 14. Classifying the plurality of speech samples into a current coded frame, dividing the coded frame into a plurality of pitch subframes, and dividing the pitch subframe into a plurality of coded subframes. Subdividing, and for the last coded subframe of each pitch subframe in the current coded frame, an initial pitch lag estimate lag ₁ and la representing a lag estimate, respectively.
evaluating g ₂ respectively, refining the pitch lag estimate lag ₀ of the second pitch subframe in the preceding coded frame, linearly interpolating lag ₁ , lag ₂ , and lag ₀ , said 14. The method of claim 13, further comprising evaluating pitch lag values for encoded subframes and further refining the interpolated pitch lag for each encoded subframe. 15. The method of claim 13, further comprising the step of down-sampling the speech samples to down-sampled values for a schematic representation with a small number of samples. 16. The method of claim 15, further comprising correcting the initial pitch lag value according to an equation (Lag _scaled = number of voice samples / downsampled value). 17. A step of receiving the speech samples, a step of processing the refined pitch lag value to reproduce the input speech as encoded speech, and a step of outputting the encoded speech. 14. The system of claim 13, further comprising: 18. A speech coding method for reproducing and coding an input speech, wherein the speech coding device induces a linear predictive coding (LPC) parameter and a speech reproduction to generate a speech. And a new codebook that represents a pseudo-random signal forming a plurality of vectors, the speech coding method comprising: receiving and processing the input speech; and processing the input speech. The step of processing comprises: determining a speech coded frame in the input speech; subdividing the coded frame into a plurality of pitch subframes; and N speech samples. Defining a pitch analysis window, the pitch analysis window extending over the pitch subframe, The processing step is a step of roughly evaluating an initial pitch lag value for each pitch subframe, and an initial pitch lag evaluation for each pitch subframe is a lag evaluation for the last coding subframe of each pitch subframe. To divide each pitch subframe into a plurality of coded subframes, interpolating the evaluated pitch lag value between the pitch subframes, and performing pitch lag evaluation for each coded subframe. Determining the input speech and refining the linearly interpolated lag value, the method further comprising: reproducing the input speech and outputting the reproduced speech according to the refined pitch lag value. Speech coding method for coding. 19. The step of processing comprises downsampling a value X to represent a small number of samples.
19. The apparatus of claim 18, further comprising the steps of down-sampling the N speech samples and correcting the pitch lag value such that the corrected lag value Lag _scaled = N / X. 20. The N speech samples have the formula N = R *.
Further comprising sampling the input speech at a sampling rate R, as determined according to X,
The method according to claim 18.