JPH07212239A - Method and device for quantizing vector-wise line spectrum frequency - Google Patents

Abstract

PURPOSE: To provide a strong and consistent device to allow high-performance vector-quantization of a short term parameter expressed as a line spectrum frequency between various speakers and handsets. CONSTITUTION: This device is provided with plural specified code books 11-14 of quantized line spectrum frequency vectors, means for retrieving each specified code book for a candidate line spectrum frequency vector similar to the non- quantized line spectrum frequency vector of an input, means 15-18 for calculating distortion for each candidate vector by using the non-quantized vector of the input, and means 19 for selecting the quantized line spectrum frequency vector from among the candidate vectors from each code book by using the calculated distortion for each candidate vector.

Description

Detailed Description of the Invention

[0001]

FIELD OF THE INVENTION The present invention relates to digital voice communication systems, particularly those which are efficient in the bits used, effective in performance between the speaker and handset, moderate in complexity, effective and simple. The present invention relates to a linear spectral frequency vector quantizer for code-excited linear prediction (CELP) speech coder with a built-in transmission error detection structure.

[0002]

2. Description of the Related Art Such a device is usually called a "codec" as a coder / decoder. The present invention is particularly suitable for digital cell networks, but is effectively used in production lines that require speech compression for speech. North American cell communication systems include current analog frequency modulation (FM) forms to digital systems. Excited linear prediction (VSELP) speech coder with sufficient speed 8.0 Kbps vector sum, convolutional coded for error protection, differential quadrature phase shift key (QP
Standard schemes using SK) modulation, time division, multiple access (TDMA) schemes have been applied by the Telecommunications Industry Association (TIA). This is thought to triple the communication capacity of the cell system. To further increase capacity by a factor of two, the TIA begins the process of evaluation and the process of selecting a codec of substantially half speed. Half-speed codec with error protection for TIA technology evaluation is 6.4
It has an overall bit rate of Kbps and a frame rate of 40 ms.
Limited to small size. Codec is believed to have a voice quality comparable to a sufficient speed standard over a wide range of conditions. These conditions include various speakers, handset effects, background noise, and channel conditions.

Codebook Excited Linear Prediction (CELP) is a technique for low speed speech coding. The basic technique consists of searching a codebook of randomly distributed excitation vectors, which are input (when filtered through a pitch and linear predictive coding (LPC) short term synthesis filter). Generate an output sequence that is closest to the sequence. To do this, all candidate excitation vectors in the codebook have to be filtered by both the pitch and LPC synthesis filters to produce a candidate output sequence that can be compared to the input sequence. I won't. This makes CELP a very computationally robust algorithm, with a typical codebook of 1024.
With 40 entries each. Furthermore, a perceptual error weighting filter is usually used, which adds to the computational load. Although high speed digital signal processors help to execute very complex algorithms such as CELP in real time, the low quality problem at low bit rates persists. Voice quality must be comparable to the 8.0 Kbps digital cell standard in order to have a codec built into the communication device.

[0004]

There are numerous displays of short term predictor parameters. Commonly used is a set of line frequency spectra. Quantization of line spectral frequencies has been the subject of some research.
Both scalar and vector quantizers are designed for this purpose. Scalar quantizers typically encode 10 line spectral frequencies to meet other objectives such as good robust performance between the speaker and handset, moderate complexity, built-in error detection capability. You need 36-40 bits to do. Efficiency on the required bits is therefore sacrificed. Vector quantizers, on the other hand, achieve efficiency in terms of bits, but with built-in error
Lack of detection capability, costly for speaker or handset dependent performance, often at the expense of high complexity.

Therefore, it is an object of the present invention to provide a robust, consistent line spectrum that provides good quantization performance of short term parameters expressed as line spectrum frequencies between different speakers and handsets. A frequency vector quantizer is provided.

Another object of the present invention is to use a minimal number of bits and to provide an efficient line spectrum frequency vector quantum with built-in error detection capability to overcome transmission errors with moderate complexity. It is to provide an oxidization device.

[0007]

SUMMARY OF THE INVENTION According to the present invention, a line spectrum that classifies unquantized line spectrum frequencies into four categories using different vector quantization tables for each category. A final choice of frequency vector quantizer and the optimal category to achieve robust performance between the speaker and the handset is provided. The present invention is a division followed by a two stage limited search process that results in an aligned set of quantized line spectral frequencies "close" to an unquantized set of medium complexity within each category. Use vector quantization. An efficient and simple transmission error detection scheme at the receiver is made possible by the partitioning properties of the vector quantization and the limited search process. The present invention uses only 26 bits to encode 10 line spectral frequencies, all capable of achieving the desired objective.

[0008]

The foregoing and other objects, aspects and advantages will be understood from the following detailed description of a preferred embodiment of the invention with reference to the drawings. The subject of the invention is an improvement of the vector quantization of the speech signal. Literature (eg Shannon
Early "A Mathematical Theory of Communicatio
n ”, Bell System Technical Journal, 27, 1948
Year), the most economical coding of information requires a bit rate no greater than the entropy of the source, which rate is larger than the coding of individual samples. It is shown that this is achieved by coding the code or vector. This was done using a codebook. Only the index (ie address) of the codebook entry is transmitted for vector transmission.
The receiver has its own copy of the codebook and uses the address to recover the transmit vector. The vector quantization of the codebook is not perfect, but contains the small but descriptive vector samples that are actually encountered in the data to be encoded. Therefore, the closest matching codebook entry is selected to send the sequence of samples and the address is sent. The vector quantization method has the advantage of reducing the bit rate, but it causes signal distortion due to the mismatch between the actual vector and the selected entry in the codebook.

In the codebook structure, the short term predictive filter coefficients of a speech frame of duration 10-30 milliseconds (ms) are obtained using conventional linear predictive analysis. The 10th order model is very common. These short term 10th order model parameters are updated at intervals of 10 to 30 ms. Quantization of these parameters is usually performed in the domain where it is perceived that the spectral distortion introduced by the quantization process is minimal for a given number of bits. One such domain is the line spectral frequency domain. A valid set of line spectral frequencies is a neatly aligned set of monotonically increasing frequencies. The complexity of converting short term predictor parameters to line spectral frequencies depends on the degree of resolution required. A small loss of performance is observed using this vector quantization scheme with a resolution of 40Hz. The 10 line spectral frequencies are quantized and encoded by the vector quantizer according to the preferred embodiment of the invention using 26 bits. After quantization, the 10 quantized line spectral frequencies are transformed into short term predictor filter coefficients.

Partition vector quantization is used to reduce the size of the codebook. The 26 bits used by the vector quantizer are spectral frequency encodes.
Includes 24 bits for encoding and 2 bits for encoding the optimal category. Aligned elements x ₁ , x with ₁₀ spectral frequencies from the lowest x ₁ to the highest x _10
2 , x ₃ , x ₄ , x ₅ , x ₆ , x ₇ , x ₈ , x ₉ , x ₁₀
, The codebook representing the set of all vectors encoded in 24 bits has 2 ²⁴ entries. Three aligned subsets, referred to herein as 3-3-4 split vectors, each encoded with 8 bits, namely (x ₁ , x ₂ , x ₃ ), (x ₄ , x ₅ , X ₆ ),
By separating the vector into (x ₇ , x ₈ , x ₉ , x ₁₀ ), three codebooks with 2 ⁸ or 256 entries respectively are required, which makes the construction of the codebook very expensive. It will be.

In basic form, the vector quantization scheme used by the present invention is 4 for the following categories of unquantized line spectral frequencies (LSF):
Two separate vector quantization tables are used.

1. IRS filtered speech LSF vector 1. 2. IRS filtered non-speech LSF vector Speech LSF vector not IRS filtered 4. Non-IRS Unvoiced LSF Vector The IRS filter is a linear phase FIR (Finite Period Impulse Response) filter, which is used to model the high pass filter effect of a handset transducer and is of its magnitude. The response complies with CCITT recommendations. In the first and third categories, 3-4-3 split vector quantization is performed using 8-, 10-, and 6-bit codebooks. In the second and fourth categories, 3-3-4 division vector quantization is 7-, 8-, and 9-.
It is done using a bit codebook. Two bits are used to encode the optimal category. Thus a total of 26 bits are used to encode 10 line spectrum frequencies.

With reference to the drawings, vector quantization is the determination of a set of aligned line spectral frequencies from the vector quantization table corresponding to each category "closest" to the unquantized line spectral frequencies. Be started by. This is accomplished in a two step process. In the first stage, shown in FIG. 1, the first split vector entry is determined from the closest corresponding vector quantization table as the weighted mean square to the unquantized line spectrum frequencies. In particular, the first split vector, shown here as (/ x ₁ ), is I
Each codebook 11,1 of each codebook of RS filtered speech LSF vector, IRS filtered non-speech LSF vector, IRS unfiltered speech LSF vector, IRS unfiltered non-speech LSF vector
It will be supplied to 2,13,14. The output of each codebook is subjected to the distortion calculation of blocks 15,16,17,18, respectively. The calculation of the strain d is a mean square calculation,

[Equation 1] Here, j = 1, ..., N, and N is the number of elements in the codebook. Thus, each codebook search and distortion calculation produces a measure of distortion d, which is provided to selector 19. The index and codebook that produces the lowest distortion, ie, the closest selection, is the first partition vector (/
X). The second and third adjacent selections are also stored.

The entries of the second partition vector are searched in a similar manner to produce the best three candidates.
However, at least 1 in 9 combinations of candidates
The search is restricted so that one is an ordered set of two. This is a very loose limitation with a small penalty in terms of both performance or complexity. The entries of the third partition vector are retrieved in a similar way subject to the same restrictions. The three best candidates from this search are also stored. There are a total of 27 candidate combinations in the particular example of 3 partition vectors, each with 3 partition candidates, but the restriction provided is that at least one combination of candidates is an aligned set. There are M combinations, where M is less than 27.

In the second stage, shown in FIG. 2, the M combinations of the fully quantized line spectrum frequency vectors are compared to the unquantized line spectrum frequencies using the Cepstral distortion measurement and the optimal A quantized line spectrum frequency vector is determined. Only combinations that result in an aligned set of line spectral frequencies are considered in the second stage, as described above. There is at least one such combination due to the restrictions placed during the search of the second and third split vectors. The mean square error distortion measurement used in the first stage is computationally simple and therefore candidates searched using the more efficient but more computationally complex Cepstral distortion measurement in the second stage. Is appropriate to simplify the list of quantized LSF vectors of In FIG. 2, the candidate M combinations represented by blocks 21, 22, 23, 24 are input to the Cepstral distortion calculator 25.

Occasionally, the calculations are done using a double Fourier transform. In particular, the speech spectrum is the product of the excitation spectra, which for speech speech is a transformation of the speech area transfer function, which consists of a series of harmonics and is a relatively smooth frequency function. If the logarithm of the spectrum is taken, adding the excitation and transfer function elements and retransforming the logarithmized spectrum gives:

[0017]

[Equation 2] Such transformations are described in the literature (“The quef by DP Bogert
rency alanysis of time series for echoes: ceptrum,
pseudo-autocovariance, cross-cepstrum, and saphe cra
cking ”, Proc. Symp., Time Series Analysis, 209
Pp. 243, 1963) (by reversing the first four letters of "spectrum").
m) ”.

The calculation of the double Fourier transform of the cepstrum is computationally complex. The preferred embodiment of the present invention therefore uses a calculation in the form of a cepstrum distance d _cep and a mean square calculation similar to that performed by the distortion calculation of FIG. In particular, septum strain calculator 25 performs the following calculations.

[0019]

[Equation 3] Where c _i is the reference (“Distance Measures for Speech Processing” by AH Gray, Jr. and JD Markel,
IEEE Transactions on Acoustics, Speech, and Signal
Processing, ASSP-24 Volume, No.5, October 1976, 380
The cepstrum coefficient, as described in (Page).
The output of block 25 is the combination of candidates with the least distortion, represented here as index I, index J, index K. Therefore, in the final stage shown in FIG. 2, the optimal category-is determined such that the optimal quantized LSF vector has the lowest cepstral distortion when compared to the original unquantized LSF vector. . The two goals of good performance and medium complexity are met by this two-stage method.

The fully quantized LSF vector is obtained from the optimal index corresponding to each of the three divided vectors by concatenating the corresponding vector entries. Such vectors are guaranteed to be ordered. Quantized LSF reconstructed at receiver
If the vectors are not aligned, three optimal vector quantization index transmission errors are detected. This simple error detection capability of this vector quantization scheme is enabled due to the limitations placed during the search period, and is made effective due to its three-part nature.

In the preferred embodiment of the invention, the vector quantization table is designed using a variation of the standard LBG algorithm using a large database of 50 or more speakers. . The average logarithmic spectral distortion over all frequencies from 0-4 Khz measured using this vector quantisation method is the IRS filtered and unirs filtered speech database. −
It is 1.3 Db over both tabes.

FIG. 3 is a block diagram showing specific settings according to the preferred embodiment of the present invention. 3-4-3 Split Band Quantization of LSF as Split Vector Quantization 31 Speech Speech, IRS Filtered, and IR
Used for unfiltered speech LSF vectors. The LSF split band portion 32 as a 3-3-4 split vector quantisation is used for non-speech speech or IRA filtered and non-IRS filtered non-speech LSF vectors.

The split band portion 31 undergoes the codebook search process shown in FIG. 4A, while the split band portion 32
Undergoes the codebook search shown at B in FIG.
First, referring to FIG. 4A, the first division vector w ₁ ,
w ₂ and w ₃ are searched in the codebook 41 to generate the combination candidates w ₁ ^* , w ₂ ^* and w ₃ ^* , and the second division vectors w ₄ , w ₅ , w ₆ and w ₇ are set. Matching candidate w ₄ ^* , w
₅ ^*, w ₆ ^*, co to produce a w ₇ ^* - codebooks 42
, And the third partition vectors w ₈ , w ₉ , w ₁₀ are combined candidate w as described normally with reference to FIG.
₈ ^*, w ₉ ^*, co-in order to generate the w ₁₀ ^* - codebook 43
It is searched by. The combinatorial candidates undergo a cepstral distance calculation at block 44 to generate a cepstral distance δ _cep for the speech portion. Similarly, in FIG. 4B, the first division vectors w ₁ , w ₂ , w ₃ are combination candidates w ₁ ^** , w ₂
^** , * w ₃ ^** (the superscript ** is used to distinguish from the combination candidates generated by the codebook 41) are searched in the codebook 45 and the second The division vectors w ₄ , w ₅ , and w ₆ are combination candidates w ₄ ^** ,
w ₅ ^**, for co generating a w ₆ ^** - is searched codebook 46, a third split vector _{w 7, w 8, w 9} , w 10 are combinatorial candidates w ₇ ^**, w ₈ ^** , W ₉ ^** , w ₁₀ ^** are searched in codebook 47. The combinatorial candidates are subjected to a cepstral distance calculation at block 48 to produce a cepstral distance * δ _cep for the non-voice portion.

Although the present invention has been described with respect to a single preferred embodiment, it will be appreciated by those skilled in the art that the present invention may be modified within the scope of the appended claims.

[Brief description of drawings]

FIG. 1 is a functional block diagram illustrating a codebook search process at a transmitter for selecting a candidate index for a transmitted split vector.

FIG. 2 is a functional block diagram illustrating a limited selection process for selecting the lowest distortion split vector sequence for transmission.

FIG. 3 is a block diagram showing a preferred configuration of the invention using 3-4-3 and 3-3-4 split vector quantization of speech and non-speech speech portions.

FIG. 4 is a block diagram showing a search process of 3-4-3 voice speech partial search, and 3-3-4 non-voice speech.
The block diagram which showed the search process of the H part.

 ─────────────────────────────────────────────────── --Continued Front Page (72) Inventor Kumar Swaminathan, State of Maryland, USA 20879, Gaithersburg, No. 202, Lost Knife Circle 18211

Claims (2)

[Claims] 1. Receiving an unquantized line spectral frequency vector and searching a plurality of specific respective codebooks for candidate quantized line spectral frequency vectors similar to the unquantized vector, Compute the distortion measure for each candidate vector using the unquantized vector and the quantized line spectrum frequency vector from among the candidate vectors from each codebook using the calculated distortion measure for each candidate vector. A method for quantizing a line spectrum frequency vector in a digital communication system, characterized in that. 2. A plurality of specialized codebooks of quantized line spectral frequency vectors and each specialized line spectral frequency vector similar to the input unquantized line spectral frequency vector. Means for searching the codebook, means for calculating the distortion measure for each candidate vector using the unquantized vector of the input, and means for calculating the distortion measure for each candidate vector from each codebook. And a means for selecting a quantized line spectrum frequency vector from the candidate vectors.