KR100220783B1 - Speech quantization and error correction method - Google Patents

Abstract

The present invention improves the quantization of the spectral amplitude by reducing the redundancy included in the spectral amplitude. The prediction of the spectral amplitude of the current segment from the spectral amplitudes of the previous segment is adjusted to indicate some variation in the fundamental frequency between the two segments. The spectral amplitude prediction errors are divided into a fixed number of blocks containing approximately the same number of elements. A predictive error block average (PRBA) vector is formed; Each element of the PRBA is equal to the average of the prediction errors in one block. PRBA vectors are quantized or transformed into discrete cosine transforms (DCT) to be scalar quantized. The detection effect of beat errors is reduced by smoothing voiced / unvoiced crystals. The estimation of the error rate is made by locally averaging the number of correctable bit errors in each segment. If the error rate is greater than the threshold, the high energy spectral amplitudes are declared voiced.

Description

[Name of invention]

Speech quantization and error correction method

Detailed description of the invention

The present invention relates to a method of preserving sound quality and quantizing speech while bit errors are present.

The following publications are related to the present invention. JL Flanagan's "Speech Analysis, Synthesis and Perception" (springer-verlag, 1972, pp. 378-386) Discloses a phase vocoder frequency based food analysis-synthesis system based on a speech analysis-synthesis system, "Speech Transformations Based on a Sinusoidal Representation" by Quatieri et al. (IEEE TASSP, Vol, ASSP 34, NO.6, Dec. 1986, pp. 1449-1986) Discloses an analysis-synthesis technique based on sinusoidal representations, Griffin's "Multiband Excitation Vocoder" (Ph.D. Thesis, MIT, 1987) Is disclosed for an 8000bps multiband excitation voice coder, Griffin et al. "A High Quality 9.6 Kbps Speech Coding System" (Proc. ICASSP 86, pp. 125-128, Tokyo, Japan, April 13-20, 1986) Discloses a 9600bps multiband excitation voice coder, Griffin's "A New Model-Based Speech Analysis / Synthesis System" (Proc. ICASSP 85. pp.513-516, Tampa, FL. March 26-29, 1985) Discloses a multi-band excitation speech model, Hardwick's "A 4.8 kbps Multi-Band Excitation Speech Coder" (SM Thesis, MIT, May 1988) Discloses a 4800bps multiband excitation voice coder, McAuly et al., "Mid-rate Coding Based on a Sinusoidal Representation of Speech" (Proc. ICASSP 90, pp. 465-468, Albeguerge, NM. April 3-6, 1990) Discloses error correction of a slow voice coder. Campbell et al. "The New 4800 bps Voice Coding Standard" (Mil Speech Tech Conference, Nov. 1989) Discloses error correction for low speed voice coders. "CELP Coding for Land Mobile Radio Applications (CELP)" by Compbell et al. (Proc. ICASSP 90, pp. 465-468, Albeguerge, NM. April 3-6, 1990) Discloses error correction of a slow voice coder. Levesque et al. "Error Control Techniques for Digital Communication" (Wiley, 1985, pp.157-170) Discloses error correction in general. Jayant et al. "Digital Coding of Waveform" (Prentice-Hall, 1984) Discloses overall quantization, Makhoul et al., "Vector Quantization in Speech Coding" (Proc. IEEE, 1985, pp. 1551-1588) Is disclosed throughout vector quantization, Jayant et al. "Adaptive Postifiltering of 16 kb / s-ADPCM Speech" (Proc. ICASSP 86, pp. 829-832, Tokyo, Japan, April 13-20, 1986) Discloses the application post filtering of speech. The contents disclosed in these published documents are included as part of this specification.

The problem of coding speech (compressing it into small bits) is of interest in many papers, as it has many applications. Voice coders (vocoders) that are widely studied and used in practice are based on the basic model of speech. For example, Linear Prediction Vocoder, Homomorphic Vocoder and Channel Vocoder are this type of vocoder. In such vocoder, speech is modeled in short time units according to the response of a linear system excited by a periodic series of impulses for voiced speech or random noise for unvoiced speech.

For this type of vocoder, speech is analyzed by segmenting the speech using a window such as a Hamming window. Then, for each segment of speech, the excitation parameter and system parameters are estimated and quantized. The excitation parameter includes voiced / unvoiced determination and pitch interval.

The system parameter includes the impulse response or spectral envelope of the system.

To reproduce the speech, the quantized excitation parameter is used to synthesize an excitation signal containing a periodic series of impulses in the voiced sound interval or random noise in the unvoiced sound interval. The excitation signal is then filtered using quantization system parameters.

Vocoders based on this basic speech model are very successful in producing readable speech, but they have not been successful in producing quality speech. As a result, the vocoder has not been widely used to code high quality speech. The poor quality of reproduced speech is partly due to incorrect estimation of the model parameters and also due to limitations of the speech model.

A new speech model, also called the Multi-Band Excitation (MBE) speech model, was developed by Griffin and Lim in 1984. Voice coders based on the new voice model were developed in 1986 by Griffin and Lim, and demonstrated that the coders were able to produce high quality voice at speeds above 8000 bps (bits per second).

Since then, the 4800 bps voice coder developed by Hardwick and Lim has been able to produce high quality voices. The 4800 bps voice coder uses a more sophisticated quantization technique to obtain a sound quality similar to that obtained by the MBE voice coder of 8000 bps at 4800 bps.

The 4800 bps MBE speech coder estimated MBE speech model parameters using an MBE analysis / synthesis system and synthesized speech from the estimated MBE speech model parameters.

A discrete speech signal represented by s (n) is obtained by sampling an analog speech signal. This is usually 8 It is performed at a sampling rate of. It is also possible to easily apply different sampling rates through simple changes of various system parameters. The system windows (n) The windowed signal s _ω (n) is obtained by multiplying (e.g., a Hamming window or Kaiser window) by s (n) to divide the discrete speech signal into small overlapping segments or segments.

Each speech segment is then analyzed to obtain a set of MBE speech model parameters that determine the characteristics of the segment. The MBE speech model parameter is a fundamental frequency ( Pitch period), a series of voiced / unvoiced judgments, a series of spectral amplitudes, and a series of spectral phases (which may or may not be included). The model parameters are then quantized using a certain number of bits for each segment. The resulting bits can then be used to reproduce the speech signal, which is performed by first reproducing the MBE model parameters from the bits and then synthesizing the speech from the model parameters. A block diagram of a typical MBE voice coder is shown in FIG.

The 4800 bps MBE voice coder required more sophisticated techniques to quantize the spectral amplitude. The number of bits that can be used to quantize the spectral amplitudes for each speech segment varies between 50 and 125 bits. In addition, a quantization technique has been devised that can efficiently represent all spectral amplitudes having a valid number of bits in each segment.

Although the spectral amplitude quantization technique is designed for use in MBE speech coders, the quantization technique is equally useful for many other speech coding techniques, such as Sinusoidal Transform Coder and Harmonic Coder. For a special speech segment, L indicates the number of spectral amplitudes of that segment. L value is the fundamental frequency ( ₀ ) is derived according to the following relationship.

Where 0 β 1.0 determines the voice bandwidth for sampling rate / 2.

The function [x] cited in equation (1) is equal to the maximum integer less than or equal to x. L spectral amplitudes are 1 One It is expressed as M₁ for L, where M₁ is the lowest frequency spectral amplitude and M₁ is the highest frequency spectral amplitude.

The spectral amplitudes for the current speech segment are quantized by calculating a series of prediction residuals representing the amount of change in the spectral amplitudes between the current speech segment and the past speech segment.

1 if L ⁰ represents the number of spectral amplitudes of the current speech segment and L ^-1 represents the number of spectral amplitudes of the past speech segment. One The prediction error Τ _ℓ for L ⁰ is given by

Where Μ ⁰ _ℓ represents the spectral amplitude of the current speech segment Denotes the quantized spectral amplitude of the past segment. a constant Is usually 7, but 0 Any value within the range of 1 can be used.

The prediction error is divided into blocks of k elements, where k is usually 4 k It is in the 12 range. If L cannot be divided evenly by k, then the highest frequency block contains a smaller number of elements than k. This is shown in Figure 2, where in Figure 2 L 34 and k 8.

Each prediction error block is then transformed using a Discrete Cosine Transform (DCT), where the DCT is defined as follows.

The length J of each transform for each block is equal to the number of elements in the block. Therefore, all blocks except the highest frequency block are converted to a DCT of length k, while the length of the DCT for the highest frequency block is less than or equal to k. Since the DCT is an invertible transform, the L DCT coefficient completely specifies the spectral amplitude prediction errors for the current segment.

The total number of bits that can be used to quantize the spectral amplitudes is divided between the DCT coefficients according to a bit allocation rule. The rule is provided for low frequency blocks that are more sensitive to sensitive than high frequency blocks that are more sensitive to sensitive. In addition, the bit allocation rule divides bits in one block into DCT coefficients according to relative long-term changes of the DCT coefficients. This method makes the bit allocation suitable for speech recognition characteristics and DCT quantization characteristics.

Each DCT coefficient is quantized using the number of bits specified by the bit allocation rule. Normally, equal quantization is used, but non-uniform or vector quantization may also be used. The step size for each quantizer is determined from the long term variance of the DCT coefficients and the number of bits used to quantize each coefficient. Table 1 below shows a typical change in step size as a function of the number of bits for a long term variance equal to σ ² .

Once each DCT coefficient has been quantized using the number of bits specified by the bit allocation rules, a binary representation can be transmitted, stored, etc., depending on the application. The spectral amplitudes can be reproduced from the binary representation, which first reproduces the quantized DCT coefficients for each block, performs an inverse DCT on each block yarn, and then uses the inverse of equation (2) to segment the past. By combining with the quantized spectral amplitudes of. The reverse DCT is given by

Where the length J for each block is selected as the number of elements in the block, and α (j) is given by

One potential problem with the 4800 bps MBE speech coder is that the recognition of the reproduced speech may be greatly reduced when bit errors are added to the binary representation of MBE model parameters. In many voice coder applications, because bit errors exist, a robust voice coder must be able to correct, detect and / or allow bit errors. One technique that has been found to be very successful is to use an error correction code of the binary representation of the model parameters.

The error correction code corrects for infrequent bit errors and allows the system to evaluate the error rate. The evaluation of the error rate can then be used to properly process the model parameters to reduce the effect of any remaining bit errors.

Typically, the error rate is evaluated by counting the number of errors corrected (or detected) by the error correction code of the current segment, and this information is used to update the current estimate of the error rate. For example, each segment contains a (23,12) Golay code that can correct three errors out of 23 bits, and ε _T is the number of errors (0-3) corrected in the current segment. In the case of displaying, the current evaluation of the error rate ε _R is updated according to the following equation.

Where β is 0, which controls the adaptability of ε _R. β Constant in the range of 1.

When an error correction code or error detection code is used, the bits representing the voice model parameter are switched to another bit set that is more robust to bit error.

The use of the error correction or detection code typically increases the number of bits that must be transmitted and stored. The number of extra bits that must be transmitted is usually related to the robustness of the error correction or detection code. In most applications, it is desirable to minimize the total number of bits transmitted and stored. Error correction or detection codes should be chosen to minimize overall system operation.

Another problem with this class of speech coding systems is that the quality of the synthesized speech is degraded due to the limitations of the estimation of the speech model parameters. Subsequent quantization of model parameters exacerbates quality degradation. This degradation can make the synthesized voice sound like an echoed voice or a closed voice. In addition, background noise or other artifacts may be issued that were not originally present in voice. This type of speech degradation occurs even when there is no bit error in the speech data. However, bit errors can exacerbate this problem.

Speech coding systems typically optimize parameter estimators and parameter quantizers to minimize this problem. Other systems try to solve the problem by post-filtering.

In post filtering, the output speech is filtered in the time domain with an adaptive all-pole filter that sharpens the format peak. The method does not provide fine control over the spectral enhancement process and is computationally complicated and inefficient for frequency domain speech coders.

The invention disclosed herein provides many other coding techniques, such as linear predictive speech coders, channel vocoders, homomorphic vocoders, sinusoidal transform coders, multiband excitation speech coders and improved multiband excitation (IMBE) speech coders. Applies to, but is not limited to. To illustrate the invention in detail, a recently standardized 6.4 Kbps IMBE voice coder may be used as part of the INMARSAT-M satellite communications system. The coder uses a robust speech model, also known as a multi-band excitation (MBE) speech model.

Effective methods for quantifying the MBE model parameters have been developed. The methods can quantize the model parameters at virtually any bit rate above 2 Kbps. The 6.4 Kbps IMBE voice coder used in the INMARSAT_M satellite communication system is 50 Move the frame rate. Therefore 128 bits per frame can be used. 45 bits out of the 128 bits are reserved for forward error correction.

The 83 bits remaining in the frame are used to localize the MBE model parameters, which are the fundamental frequencies. ₀ , 1 k V / UV Determination for K υ _k and 1 One The spectral amplitude M₁ parameter for L is included.

The K and L values change with the fundamental frequency of each frame. The 83 valid bits are divided between the model parameters as shown in Table 2 below.

The fundamental frequency is quantized by converting the fundamental frequency into its equivalent pitch section using equation (7).

The value of P ₀ is usually 8 20 for sampling rate P ₀ Limited to 120 ranges. In a 6.4 Kbps IMBE system, the parameters are quantized evenly using 8 bits and a step size of 5. This corresponds to a pitch accuracy of 1/2 sample.

The KV / UV Decision Parameter Binary Gad. Therefore they can be coded using one bit for one decision. The 6.4 Kbps system uses up to 12 decision parameters and each frequency bandwidth is 3 ₀ . The highest frequency is 3.8 Adjusted to include frequencies up to

The spectral amplitudes are normalized by forming a series of prediction errors. Each prediction error is the difference between the logarithm of the spectral amplitude for the current frame and the logarithm of the spectral amplitude representing the same frequency of the past speech frame. The spectral amplitude prediction error is then divided into six blocks containing each approximately equal number of prediction errors.

Each of the six blocks is transformed by a Discrete Cosine Transform (DCT), and the D.C coefficients of each of the six blocks are combined to form a single 6-element Prediction Residual Block Average (PRBA) vector. The mean is subtracted from the PRBA vector and quantized using a 6 bit heterogeneous quantizer. The zero mean PRBA vector is then vector quantized using a 10 bit vector quantizer. The 10-bit PRBA codebook was designed using a K-means clustering algorithm on a large training set containing zero mean PRBA vectors from various speech materials.

Higher order DCT coefficients not included in the PRBA vector are quantized with a scalar uniform quantizer using the '59 -K 'residual bits. Bit allocation and quantizer step size are based on long term variance of higher order DCT coefficients.

The quantization technique has several advantages.

First, the technique provides very good fidelity with a few bits and maintains the fidelity even if the range of L changes. The computational requirements of the technique are also within the limits required for real-time implementation using a single DSP, such as the AT & T DSP32C. Finally, the quantization technique separates the spectral amplitudes into some elements that are sensitive to bit errors, such as the average of a PRBA vector, and many other elements that are very insensitive to bit errors. Therefore, it is possible to effectively use forward error correction by providing a high degree of protection for some elements that are sensitive to bit errors and a low level of protection for the remaining elements. This will be explained below.

According to a first aspect of the present invention there is provided an improved method for forming a predicted spectral amplitude. The method is based on interpolating the spectral amplitude of the past segment to estimate the spectral amplitude of the past segment at the frequency of the current segment. This new technique makes the variance of the prediction error smaller by correcting the shift in the frequency of the intersegment spectral amplitude, and thus can be quantized with smaller distortion for a given number of bits. In a preferred embodiment, the frequency of the spectral amplitudes is a fundamental frequency and a multiple of this fundamental frequency.

According to a second aspect of the present invention, there is provided an improved method for dividing a prediction error into blocks. Instead of dividing the prediction errors into a variable number of blocks after fixing the length of each block, the prediction errors are divided into a predetermined number of blocks and the block size is varied according to segments. In a preferred embodiment, six blocks are used for all segments, and the number of prediction errors of the low frequency block is not greater than the number of prediction errors of the high frequency block, and the difference between the number of elements of the highest frequency block and the number of elements of the lowest frequency block Is less than or equal to 1 Since this new technique is closer to the speech characteristic, the new technique allows the prediction error to be quantized with little distortion for a given number of bits. In addition, the technique can be easily used with vector quantization to further improve the quantization of the spectral amplitudes.

According to a third aspect of the present invention, there is provided an improved method for quantizing prediction errors. The prediction errors are systematically classified in blocks to determine the average of the prediction errors in each block. The average of all blocks is systematically classified into a Prediction Residual Block Average (PRBA) vector, and the PRBA vector is encoded. In a preferred embodiment, the average of the prediction errors adds spectral amplitude errors into the block and divides the number of prediction errors into the block or calculates the DCTs of the spectral prediction errors within one block to obtain the first coefficient of DCT as the average. Is obtained. The PRBA vector is preferably encoded using one of the following two methods.

(1) performing a transform such as DCT of a PRBA vector and scalar quantizing the transform coefficients.

(2) A method of vector quantization of the PRBA vector.

Vector quantization is preferably performed by determining the mean of the PRBA vectors, quantizing the mean using scalar quantization, and quantizing the zero mean PRBA vector using vector quantization with a zero mean codebook. An advantage of the present invention is that it is possible to quantize the prediction errors while reducing the distortion for a given number of bits.

According to a fourth aspect of the present invention, there is provided an improved method for making voiced / unvoiced decisions when there is a higher bit error rate. The bit error rate is evaluated for the current speech segment and compared with a predetermined error rate threshold.

If the evaluated bit error rate exceeds the error rate threshold, then voiced / unvoiced determinations for spectral amplitudes exceeding a predetermined energy threshold make the decision that all voices are voiced for the current segment. This reduces the effect of recognizing bit errors. The distortion produced by changing from voiced to unvoiced is reduced.

According to a fifth aspect of the invention, there is provided an improved method for error correction (or error detection) coding of speech model parameters. The new method uses at least two types of error correction coding to code the quantization model parameters. The first type of coding adds a larger number of additional bits than the second type of coding, which is used for parameter groups that are more sensitive to bit errors. Another type of error correction coding is used for the second parameter group which is less sensitive to bit errors than the first type. Comparing this method with those currently used, the new method improves the quality of synthesized speech when bit errors are present, while reducing the amount of additional error correction or detection bits that must be added. In a preferred embodiment, the other type of error correction includes a Golay code and a Hamming code.

According to a sixth aspect of the present invention, there is provided another method for improving the quality of synthesized speech when bit errors are present. The error rate is evaluated from error correction coding and one or more model parameters from past segments are repeated in the current segment if the error rate for parameters exceeds a certain level. In a preferred embodiment, all model parameters are repeated.

According to a seventh aspect of the present invention, there is provided a new method for reducing distortion caused by evaluation and quantization of model parameters. The new method utilizes parking area representing the spectral envelope parameters to increase the sensibly important spectral area and reduce the sensitively less significant area. As a result, the quality degradation of the synthesized voice is reduced. The smoothed spectral envelope of the segment is generated by smoothing the spectral envelope, and the increased spectral envelope increases a certain frequency range for the spectral envelope with a larger amplitude than the smoothed envelope and has a smaller amplitude than the smoothed envelope. It is generated by reducing any frequency range for the envelope. In a preferred embodiment, the smoothed spectral envelope is generated by evaluating a low-order model (eg, an all-pole model) from the spectrum envelope. Compared to currently used methods, the new method is more computationally efficient for frequency domain voice coders. The new method also improves voice quality by eliminating the frequency domain suppression imposed by the time domain method.

Other features and advantages of the present invention will become more apparent from the following description of the preferred embodiments and the claims.

[Brief Description of Drawings]

1-2 is a block diagram showing a conventional speech coding method.

3 is a flow chart showing some variation in fundamental frequency where spectral amplitude prediction is a preferred embodiment of the present invention.

4 is a flowchart showing that the spectral amplitude is divided into a fixed number of blocks as a preferred embodiment of the present invention.

5 is a flowchart showing that a prediction error block average vector is formed as a preferred embodiment of the present invention.

6 is a flowchart showing that the prediction error block mean vector is vector quantized as a preferred embodiment of the present invention.

7 is a flowchart showing that a prediction error block mean vector is quantized by DCT and scalar quantization as a preferred embodiment of the present invention.

8 is a flowchart showing an encoder in which different error correction codes are used for different model parameter bits as a preferred embodiment of the present invention.

9 is a flowchart showing a decoder in which different error correction codes are used for different model parameter bits as a preferred embodiment of the present invention.

10 is a flow chart illustrating the enhanced frequency domain spectral envelope parameters as a preferred embodiment of the present invention.

[Description of Preferred Embodiments of the Invention]

In the prior art, spectral amplitude prediction errors were formed using equation (2). Such a method cannot account for any change in the fundamental frequency between the past segment and the current segment. To account for the change in the fundamental frequency, a new technique has been developed that first interpolates the spectral amplitudes of the go-go segments. This is typically done using linear interpolation. However, various other forms of interpolation may be used. The interpolated spectral amplitudes of the past segment are then resampled at a frequency point corresponding to a multiple of the fundamental frequency of the current segment. This combination of interpolation and resampling produces a series of predicted spectral amplitudes that are corrected for any intersegment changes in the fundamental frequency.

Typically, the logarithm of some predictive spectral amplitudes of base 2 is subtracted from the logarithm of the spectral amplitudes of the current segment of base 2. When linear interpolation is used to calculate the prediction spectral amplitudes, this is expressed mathematically as follows.

Where δ _ℓ is given by

here Is 0 An integer belonging to 1. Normal, 7, but of a different value May also be used. E.g May be changed according to the segment to improve performance. Parameters of Equation (9) Wow Are the fundamental frequencies of the current and past segments, respectively. If the two fundamental frequencies are the same, the new method produces prediction errors with lower variance than the conventional method. This produces prediction errors with less distortion for a given number of bits.

In another aspect of the present invention, a new method of dividing the spectral amplitude prediction errors into blocks has been developed. In the conventional method, the prediction error L from the current segment is divided into blocks consisting of K elements. Where usually K 8. Using this method, it is found that the characteristics of each block differ greatly for large and small L values. This reduces the quantization efficiency and thus increases the distortion of the spectral amplitudes.

In order to make the characteristics of each block more uniform, a new method of dividing the prediction error L into a fixed number of blocks has been devised. The length of each block is selected so that all blocks in one segment have almost the same length and the sum of the lengths of all blocks in one segment is equal to L. The number of prediction errors is divided into six blocks, where the length of each block is equal to [L / 6]. If L cannot be divided equally into six, the length of one or more high frequency blocks is increased by one so that all spectral amplitudes are included in one of the six blocks. This new method is shown in Figure 4, where six blocks are used and L 34.

In the new method, an approximation of the prediction errors included in each block is independent of L. This reduces the characteristic change of each block and provides more efficient quantization of the prediction errors.

Quantization of the prediction errors can be further improved by forming a prediction error block mean (PRBA) vector. The length of the PRBA vector is equal to the number of blocks in the current segment. The elements of the vector correspond to the average of the prediction errors in each block. Since the first DCT coefficients are equal to the mean (or DC value), a PRBA vector can be formed from the first DCT coefficients of each block. This means that there are 6 blocks in the current segment and L FIG. 5 is also shown for the case 34. The process is generalized by forming additional vectors from the second (or third, fourth, etc.) DCT coefficients of each block.

Correlation of the elements of the PRBA is very high. Therefore, several methods can be used to improve the quantization of the spectral amplitudes. One method that can be used to achieve very low distortion with a small number of bits is vector quantization. In this method, a codebook is devised that includes a number of typical PRBA vectors. The PRBA vector for the current segment is compared with each codebook vector and selects the vector with the smallest error as the quantized PRBA vector. The codebook index of the selected vector is used to form a binary representation of the PRBA vector. A method of performing vector quantization of PRBA using a cascade of a 6-bit nonuniform quantizer with respect to the average of the vectors and a 10-bit vector quantizer for information maintenance has been developed.

This method is shown in FIG. 6 for the case of a PRBA vector that always contains six elements.

Another method for quantizing PRBA vectors has also been developed. This method requires less computation and storage than the vector quantization method. In this way, the PRBA vector is first transformed into a DCT defined by equation (3). The DCT length is equal to the number of elements of the PRBA vector. The DCT coefficients are then quantized in a manner similar to that described in the prior art. Initially, a bit allocation rule is used to divide the total number of bits used to quantize the PRBA vector between the DCT coefficients. Then, using scalar quantization (uniform or non-uniform), each DCT coefficient is quantized using the number of bits specified by the bit allocation rule. This is shown in FIG. 7 for the case where the PRBA vector always contains six elements.

Various other methods can be used to effectively quantize PRBA vectors. For example, other transforms such as Discrete Fourier Transform (DFT), Fast Fourier Transform (FFT), Karhunen_Louve Transform can also be used instead of DCT. Vector quantization can also be combined with DCT or other transforms. The improvements resulting from the present invention can be used with a wide variety of dressing methods.

As another method of the present invention, a new method for reducing the recognition effect of bit error has been developed. Error correction codes are used as in the prior art to correct for rare bit errors and provide an estimate of the error rate ε _R. To reduce the recognition effect of all remaining bit errors, the new method uses error rate estimates to smooth voiced / unvoiced decisions. This is done by first comparing the threshold with the error rate, where the threshold is the error rate at which the distortion resulting from the uncorrected bit error of the voiced / unvoiced decision parameter is at a high rate.

The exact value of the threshold depends on the amount of error correction applied to the voiced / unvoiced determination, but when there is little applied error correction, the threshold is typically .003. If the estimated error rate ε _R is less than the threshold, voiced / unvoiced judgment is not confused. If ε _R exceeds the threshold, all spectral amplitudes that satisfy Equation 10 below are declared voiced.

Although the threshold is assumed to be .003 in equation (10), the method can easily be modified for other thresholds. The parameter S _E is a measure of the partial mean energy contained in the spectral amplitude. The parameter is updated as follows for each segment.

Where R ₀ is given by

The initial value of S _E is 0 SE It is set to any initial value in the range 10000.0. This parameter is set in this way in order to reduce the dependence of Equation (10) on the average signal level. In this way the new method is performed on low level signals as well as on high level signals.

The feature forms of formulas (10), (11) and (12) and the constants included in this formula can be easily changed while maintaining the original components of the new method. The main component of the new method is to initially use the error rate estimate to determine whether smoothing of the voiced / non-voiced determination parameters is required. If smoothing is required, voiced / unvoiced judgment parameters are disturbed such that all high energy spectral amplitudes are declared voiced. Thus, no high-energy voiced sound is converted from the voiceless voice to the voiced voice or the voiced voice, so that the recognition of the reproduced voice is improved when there is a bit error.

The present invention divides the quantized speech model parameter bits into three or more groups according to their sensitivity to their bit errors, and then uses different error correction or detection codes for each group. Typically, groups of data bits that are determined to be most sensitive to bit errors are protected using highly effective error correction codes. Less effective error correction or detection codes that require little additional bits are used to protect less sensitive data bits. The new method allows the amount of error correction or detection given in each group to match the sensitivity to bit errors. Compared with the prior art, the method has the advantage of reducing the query degradation caused by the bit error and of course reducing the number of bits requiring forward error correction.

Selecting the error correction or detection code used depends on the bit error statistics and the desired bit rate of the transmission medium or storage medium. The most sensitive group of bits is typically protected by an effective error correction code such as a Hamming code, BCH code, Golay code or Reed-Slolmon code. Groups of less sensitive data bits may use the error correction code and may use the error detection code. The least sensitive group may use error correction or detection codes or no form of error correction or detection. The present invention described above uses an error correction code and an error detection code that are well suited for a 6.4 Kbps IMBE voice coder for satellite communications.

In the 6.4 Kbps IMBE voice coder, where the INMARSAT_M (International Maritime Satellite Organization) satellite communication system was standard, 45 bits per frame preserved for forward error correction can correct up to three errors [23, 12] Golay code It is split between [15, 11] Hamming codes that can correct for single errors and parity bits. The six most significant bits from the fundamental frequency and the three most significant bits from the average of the PRBA vectors are first combined with the three parity check bits and coded with the [23,12] Golay code.

The second Golay code is used to code nine maximum sensitive bits from three maximum effective non-high order DCT coefficients from the PRBA vector. All well bits except seven minimum sensitive bits are then coded into five [15, 11] Hamming codes. The seven least significant bits are not protected by the error correction code.

Prior to transmission, 128 bits representing a particular speech segment are interleaved so that any two bits are separated from the same code word by at least five bits. Due to this characteristic, the burst error occurring in a short time is distributed to several different code words, which increases the possibility of the error being corrected.

At the decoder, the received bits are sent through the Golay and Hamming decoders to remove all bit errors from the data bits. The three parity check bits are examined and the received bits are used to reproduce the MBE model parameters for the current frame when no bit els that cannot be corrected are detected. If a bit error that cannot be corrected is detected, the receive bit for the current frame is ignored and the model parameters from the past frame are repeated for the current frame.

Using frame repetition has been found to improve speech recognition in the presence of bit errors. The present invention examines the received bits of each frame to determine whether the current frame contains many uncorrectable bit errors. One way to detect uncorrectable bit errors is to check the extra parity bits inserted in the data. The present invention also determines whether a large number of burst bit errors have occurred by comparing a local estimate of the error rate with the number of predictable bit errors. If the number of correctable bit errors is actually greater than the local estimate of the error rate, frame repetition is performed. In addition, the present invention checks each frame for an invalid bit sequence (ie, a group of bits that the encoder never transmitted). If an invalid bit sequence is detected, frame repetition is performed.

Golay and Hamming decoders also provide information about the number of correctable bit errors in the data. This information is used by the decoder to estimate the bit error rate. The estimation of the bit error rate is used to control the adaptive smoother which improves the quality of the sensed speech when there is an uncorrectable bit error. In addition, estimation of the error rate can be used to perform frame repetition in a poor error environment.

The use of the present invention in conjunction with soft-decision coding further improves performance. Soft-decision coding uses additional information about the presence of each bit error to improve the error correction and detection capabilities of many other codes.

Since the additional information is available from the modulator of the digital communication system, it can provide improved robustness to the bit error without requiring additional bits for error protection.

The present invention utilizes a new frequency domain parameter enhancement method that improves the quality of synthesized speech. The present invention initially detects important areas of the speech spectrum. The present invention then increases the width of the frequency domain which is sensitive to other frequency domains. A preferred method for performing frequency domain parameter augmentation is to smooth the spectral envelope that estimates the spectra of one form.

The spectrum can be smoothed by providing a lower order model, such as an all-pole model, a cepstral model or a polynomial model, to the spectral envelope. The smoothed spectral envelope is compared to the non- smoothed spectral envelope, and the sensitively sensitive frequency region is considered to be the region where the non- smoothed spectral envelope has more energy than the smoothed spectral envelope.

Likewise, an area where the unsmoothed spectral envelope has less energy than the smoothed spectral envelope is considered to be a less sensitive area. Parameter augmentation is performed by increasing the amplitude of the sensitively sensitive frequency domain and decreasing the amplitude of the less sensitively sensitive frequency domain. This new augmentation technique improves speech quality by eliminating or reducing many artifacts introduced during speech parameter estimation and quantization. The new method also improves clarity of speech by sharpening sensitively important speech formants.

The IMBE speech decoder provides a first order all-pole model to the spectral envelope for each frame. This is done by estimating the correlation parameters R ₀ and R ₁ from the decoded model parameters according to the following equation.

Where 1 One M₁ for L is the decoded spectral amplitude for the current frame, ₀ is the decoded fundamental frequency for the current frame. Correlation parameters R ₀ and R ₁ may be used to estimate the first order all-pole model. The model is a frequency corresponding to the spectral amplitude for the current frame (i.e. 1 One K against L ₀ ) and can be used to produce a series of weights W ₁ according to the following formula:

The weight represents the ratio of the smoothed all-pole spectrum to the IMBE spectral amplitudes. These are then applied to each spectral amplitude and used to individually control the amount of parameter amplification. Such a relationship is expressed by the following equation.

Where 1 One M₁ for L is the increased spectral amplitude for the current frame.

The increased spectral amplitudes are then used for speech synthesis. The use of the augmented model parameter improves the quality of speech as compared to synthesis from non-augmented model parameters.

Claims (37)

Voice is divided into a plurality of segments, each of which represents one section of successive time periods and has a spectrum of frequencies, for each segment the frequencies at which the spectrum is sampled as a series of frequencies are mutually different according to the segment. Different, using the spectral amplitudes for one or more past segments to generate a series of predicted spectral amplitudes for the current segment, the current based on the difference between the actual spectral amplitudes for the current segment and the predicted spectral amplitude for the current segment A series of prediction errors for a segment is used in a subsequent coding operation wherein the predicted spectral amplitudes for the current segment are used to estimate the spectral amplitude of the past segment at the frequency of the current segment. And based at least in part on interpolating spectral amplitudes of past segments. Speech is divided into a plurality of segments, each segment representing one of successive time periods and having a spectrum of frequencies, for each segment the spectrum is sampled at a series of frequencies to form a series of actual spectral amplitudes. And the frequency at which the spectrum is sampled differs from the segment, producing a series of predicted spectral amplitudes for the current segment using the spectral amplitudes for one or more past segments, and the actual spectral amplitude for the current segment. And a series of prediction errors for the current segment are used for subsequent encoding operations, wherein the prediction errors for one segment are grouped into a predetermined number of blocks, the number of blocks for a particular block. The number of errors and Tube, and the blocks are coded audio method characterized in that the encoding. 3. The speech coding of claim 2, wherein the predicted spectral amplitudes for the current segment are based at least in part on interpolating the spectral amplitude of the past segment to estimate the spectral amplitude of the past segment at the frequency of the current segment. Way. Speech is divided into a plurality of segments, each segment representing one of successive time periods and having a spectrum of frequencies, for each segment the spectrum is sampled at a series of frequencies to form a series of actual spectral amplitudes. And the frequency at which the spectrum is sampled differs from the segment, producing a series of predicted spectral amplitudes for the current segment using the spectral amplitudes for one or more past segments, and the actual spectral amplitude for the current segment. And a series of prediction errors for the current segment based on the difference between the prediction spectrum amplitudes for the current segment and the current segment are used for subsequent coding operations, wherein the prediction errors for one segment are grouped into a plurality of blocks, The average of the prediction errors in each block is determined, the average of each block is grouped into a prediction error block average (PRBA) vector, and the PRBA vector is coded. 5. The speech coding method according to claim 4, wherein there is a predetermined number of blocks, and the number of blocks of disputes is independent of the number of prediction errors grouped into specific blocks. 6. The speech coding of claim 5, wherein the predicted spectral amplitudes for the current segment are based at least in part on interpolating the spectral amplitude of the past segment to estimate the spectral amplitude of the past segment at the frequency of the current segment. Way. 2. The method of claim 1, wherein the difference between the actual spectral amplitudes for the current segment and the predicted spectral amplitudes for the current segment is formed by subtracting some predicted spectral amplitudes from the actual spectral amplitudes. 2. The method of claim 1, wherein the spectral amplitudes are obtained using a multiband excitation speech model. 2. The method of claim 1, wherein only the spectral amplitudes from the most recent past segment are used to form the predicted spectral amplitudes of the current segment. 2. The method of claim 1 wherein the spectrum comprises a fundamental frequency and the series of frequencies for a given segment is a multiple of the fundamental frequency of the segment. The speech coding method of claim 2, wherein the number of blocks is six. 3. The speech coding method of claim 2, wherein the number of prediction errors of the low frequency block is not greater than the number of prediction errors of the high frequency block. 12. The method of claim 11, wherein the number of prediction errors of the low frequency block is no greater than the number of prediction errors of the high frequency block. 15. The method of claim 13, wherein a difference between the number of elements of the highest frequency block and the number of elements of the lowest frequency block is less than or equal to. 7. The method of claim 4, 5 or 6, wherein the average is calculated by summing prediction errors in the block and dividing by the prediction errors in the block. 16. The method of claim 15, wherein the average is obtained by calculating a discrete cosine transform (DCT) of spectral amplitude prediction errors in the block and using the first coefficient of the DCT as the average. 7. The method of claim 4, 5 or 6, wherein the coding of the PRBA vector comprises vector quantization of the PRBA vector. 18. The method of claim 17, wherein the vector quantization comprises: determining an average of the PRBA vectors; quantizing the average using scalar quantization; And quantizing the zero mean PRBA vector using vector quantization with a zero mean codebook. A method for synthesizing speech from a received bit stream in which each speech segment or frequency band within the segment is decoded into voiced or unvoiced speech and representing a speech segment and having a bit error, wherein the bit error rate for the current speech segment is estimated and is Compared with the error rate threshold, when the estimated bit error rate exceeds the error rate threshold, voiced / unvoiced judgments for spectral amplitudes exceeding a predetermined energy threshold are all declared voiced for the current segment. Speech synthesis method. 20. The speech synthesis method of claim 19, wherein the predetermined energy threshold is dependent on an estimate of bit error for the current segment. Speech is coded using the speech model specified by model parameters, speech is divided into time segments, model parameters are estimated and quantized for each segment, and at least some of the quantized model parameters are error corrected coding. In the speech coding method coded using, the quantized model parameter is coded using at least two types of error correcting coding, wherein the first type of coding has a larger number of additional bits than the second type of coding. And used for a quantized model parameter of the first group, wherein the quantized model parameter of the first group is more sensitive to bit errors than the quantized model parameter of the second group. 22. The method of claim 21, wherein the error correcting coding comprises a Golay code and a Hamming code. Speech is coded using a speech model that is specific to model parameters, speech is divided into time segments, parameters are estimated and quantized for each segment model, and at least some of the quantized model parameters may be error corrected coding. A speech coding method wherein a speech is synthesized from the decoded and quantized model parameters using the error correcting coding to estimate the error rate, and one or more model parameters from past segments And repeating in the current segment when the error rate exceeds a predetermined level. 22. The method of claim 21, wherein the quantized parameters are parameters associated with a multi band excitation (MBE) voice coder or an enhanced multi band excitation (IMBE) voice coder. 23. The speech encoding method according to claim 21 or 22, wherein the error rate is estimated using an error correction code. 27. The method of claim 25, wherein the one or more model parameters are smoothed over the plurality of segments based on the estimated error rate. 27. The method of claim 26, wherein the smoothed model parameters comprise voiced / unvoiced determination. 27. The method of claim 26, wherein the smoothed model parameters comprise parameters for a multi band excitation (MBE) voice coder or an enhanced multi band excitation (IMBE) voice coder. 29. The method of claim 28, wherein the value of one or more model parameters in the past segment is repeated in the current segment if the estimated error rate for the parameters exceeds a predetermined level. In a speech enhancement method in which a speech signal is divided into segments, a frequency-domain representation of the segment is determined to provide a spectral envelope of the segment, and the speech is synthesized from an augmented spectral envelope, the smoothed spectral envelope of the segment Formed by smoothing, wherein an increased spectral envelope is formed by increasing the frequency domain of the spectral envelope having a larger amplitude than the smoothed envelope and reducing the frequency domain of the spectral envelope having a smaller amplitude than the smoothed envelope. Voice coding method. 31. The method of claim 30, wherein the frequency domain representation of the spectral envelope is a series of spectral amplitude parameters of a multiband excitation (MBE) speech coder or an enhanced multiband excitation (IMBE) speech code. 32. The method of claim 30 or 31, wherein the smoothed spectral envelope is formed by estimating a lower order model from the spectral envelope. 33. The method of claim 32, wherein the lower order model is an all-pole model. 7. The method of claim 4, 5 or 6, wherein the PRBA vector is coded using a linear transform on the PRBA vector and scalar quantization of the transform coefficients. 35. The method of claim 34, wherein the linear transform comprises a discrete cosine transform. 3. The method of claim 2, wherein the spectral amplitudes are obtained using a multiband excitation speech model. 5. The method of claim 4, wherein the spectral amplitudes are obtained using a multiband excitation speech model.