Stochastic Codebook Research Articles

Unified pulse-replacement search algorithms for algebra codebooks of speech coders

The algebraic code excited linear prediction (ACELP) algorithm, because of low complexity and high quality in its analysis-by-synthesis optimisation, has been adopted by many speech coding standards. This study proposes the unified generalised pulse replacement (UPR) search algorithm for the stochastic codebook of ACELP speech coders. The proposed UPR algorithm discusses the search breadth, the order of the search direction and the update frequency based on the pulse replacement method. In addition, there are many derivative types of UPR algorithms discussed. The proposed approaches can achieve the lowest computational complexity with imperceptible degradation of the speech quality. Furthermore, the normalised degradation ratio based on the standard subjective quality measurement is proposed to fairly compare the performance. The experimental results will verify the claims.

Switching Search Method for Pulse Assignment in ITU-T G.729D

In this paper, the simplified search designs for the stochastic codebook of algebraic code excited linear prediction (ACELP) for ITU-T G.729D speech coder are proposed. By using two search rounds and limiting the search range, the computational complexity of the proposed approach is only 6.25% of the full search method recommended by G.729D. In addition, the computational complexity of proposed approach is only 59% of the global pulse replacement search method recommended by G.729.1. Simulation results show that the coded speech quality evaluated by using the standard subjective and objective quality measurements is with perceptually negligible degradation.

Generalised candidate scheme for the stochastic codebook search of scalable CELP coders

MPEG-4 CELP speech coders configured with multiple bit rates and coding layers provide SNR scalability, and hence may be operated effectively in variable bandwidth environments. A generalised candidate (GC) scheme is proposed to reduce the computational complexity of the stochastic codebook search of CELP coders. The experimental results demonstrate that the proposed GC scheme, incorporated with the verification model (VM) of the MPEG-4 CELP coder, enables a reduction of over 50% of the computational load without suffering any subjective quality degradations. The proposed GC scheme is not only suitable for the base layer and enhancement layers, but also facilitates computational scalability, and is therefore suitable for different working platforms and integrated media source services.

A 6.7 kbps vector sum excited linear prediction on TMS320C54X digital signal processor

In this paper, a 6.7-kbps vector sum excited linear prediction (VSELP) coder with less computational complexity is presented. A very efficient VSELP codebook with nine basis vectors and a heuristic K-selection method (to reduce the search space and complexity) is constructed to obtain the stochastic codebook vector. The nine basis vectors are obtained by optimizing a set of randomly generated basis vectors. During the optimization process, we have trained the basis vectors to give the system apriori knowledge of the characteristics of the input. The coder is implemented on a TMS320C541 digital signal processor. The performance is evaluated by testing the 6.7-kbps VSELP coder with different test speech data taken from different speakers. The quality of the coder is estimated by comparing the performance of the 6.7-kbps VSELP coder with an 8-kbps VSELP speech coder based on the IS-54 standards.

Code excitation linear predictive (CELP) encoder and decoder and code excitation linear predictive coding method

There is provided a code excitation linear predictive (CELP) coding or decoding apparatus in which a code vector, which is transmitted by a codebook such as a stochastic codebook, is converted adaptively in accordance with vocal tract analysis information (LPC) so that a high quality reproduction speech is obtained at a low coding rate. Further, in order to obtain a similar effect, a pulse-like excitation codebook formed of an isolated impulse is provided in addition to the adaptive excitation codebook and stochastic excitation codebook so that either the stochastic excitation codebook or the pulse-like excitation codebook is selectively used to provide a vocal tract parameter as a linear spectrum pair parameter.

Speech compressor using trellis encoding and linear prediction

A speech compressor utilizing Trellis Encoding and Linear Prediction (TELP). A TELP speech compressor provides improved signal generation and search technique for a code-excited linear prediction (CELP) speech encoder. TELP is a frame oriented coding that breaks the quantized speech signals into frames of prescribed length N and each frame into subframes of prescribed length L, which are processed as dependent units utilizing an analysis-by-synthesis approach. The approach is based on constructing the best mean square linear predicting filter and searching the best exciting sequence for the filter in order to produce synthesized speech. A trellis encoder is used instead of a stochastic code book. The Q-ary analysis of a given subframe and previous excitations is proposed for a fast vector search in an adaptive code book. It simplifies the implementation of digital speech compression.

VLSI architecture and implementation for FS1016 CELP decoder with reduced power and memory requirements

The code-excited linear predictive (CELP) coder is the most effective technique among various linear predictive coding methods for speech compression. Hence, designing a relatively low-cost and low-power CELP decoder chip for the portable systems and wireless digital communication environment becomes increasingly important. This paper presents the VLSI architecture and chip implementation for the FS1016 CELP decoder with reduced power and memory requirements. A single-chip implementation of the CELP decoder drastically reduces the cost and the size of many CELP vocoder systems. The decoder chip can achieve the following: (1) excellent accuracy results due to the accuracy studies for the finite word length, (2) power savings and high-speed operations owing to the combined advantages of pipeline and concurrent processing structures, (3) table size reducing by applying the memoryless realization for stochastic codebook and partial sums technique and (4) specification satisfying the FS1016 CELP coder. Fabricated with 0.8 μm double-metal CMOS technology, the chip contains approximately 13 000 transistors occupying a 6.1 × 6.2 mm 2 area. It has been tested to be fully functional at IMS XL-60 tester and is the first working FS1016 CELP 4.8 K decoder chip. In particular, the fixed-point accuracy studies for the CELP decoder are also addressed herein.

Pitch synchronous multi-band (PSMB) coding of speech signals

A novel speech coding algorithm, named pitch synchronous multi-band (PSMB), is proposed. The new coding algorithm uses the multi-band excitation (MBE) model to generate a representative pitch-cycle waveform (PCW) for each frame. The representative PCW of a frame is encoded by two out of three codebooks depending upon whether the frame is related or unrelated to the previous frame. When a frame is related to its previous frame, the PCW is encoded by a length-converted-excitation (LCE) codebook and a stochastic codebook. The codevectors of the LCE codebook are derived from the previous PCW. When a frame is unrelated to its previous frame, it is encoded by a bandlimited single pulse excitation (BSPE) codebook and the stochastic codebook. The new speech coder introduces a pitch-period-based coding feature. It overcomes some weaknesses existing in the improved MBE (IMBE) speech coder. The PSMB coder operating at 4 kbps outperforms the Inmarsat 4.15 kbps IMBE coder by a clear margin. Our listening tests also indicate that it is slightly better than the FS1016 4.8 kbps code excited linear predictive (CELP) coder in terms of perceptual quality. Fast search algorithms for the three codebooks used in PSMB are also developed. The fast algorithms render the new speech coder comparable to the FS1016 CELP coder, in terms of computational complexity.

Glottal source modeling for voice conversion

This paper describes recent advances in glottal source modeling for speech synthesis. In particular two procedures for modeling the glottal excitation waveform are described and applied to voice conversion. One model uses a polynomial to represent the glottal excitation waveform for one pitch period. The coefficients of the polynomial model form a vector that is used to design a glottal excitation code book with 32 entries for voiced excitation. The codebook is designed and trained using two sentences spoken by different speakers. Speech is synthesized using a quantized glottal excitation waveform for one speaker as the excitation for a glottal excitation linear predictive (GELP) synthesizer designed using tract parameters obtained from the speech of another speaker. Our implementation of the LP synthesizer is patterned after both a pitch-excited LP speech synthesizer and a code excited linear predictive (CELP) speech coder. In addition to the glottal excitation codebook, we use a stochastic codebook with 256 entries for unvoiced noise excitation. Analysis techniques are described for constructing both codebooks. The GELP synthesizer, which resynthesizes speech with high quality, provides the speech scientist with a simple speech synthesis procedure that uses established analysis techniques, that is able to reproduce all speech sounds, and yet also has an excitation model waveform that is related to the derivative of the glottal flow and the integral of the residue. Another approach uses the LF glottal volume-velocity waveform to model the characteristics of three voice types: modal, breathy, and vocal fry (creaky). We then convert a modal voice to sound like a breathy or vocal fry voice using the vocal tract characteristics for modal voice and the glottal volume-velocity waveform model for breathy and vocal fry voices as the excitation.

Dynamic partial search scheme for stochastic codebook of FS1016 CELP coder

The authors present a new classified dynamic partial search structure for the stochastic codebook of the FS1016 CELP coder to replace the fixed partial search for selecting the best excitation vector of the stochastic codebook. In the proposed scheme, the conventional one-stage stochastic codebook search is substituted with a two-stage dynamic method for reducing the computational complexity without degrading the voice quality. The establishment of this structure is based on two classifiers, one for the line spectrum pairs (LSP) of the input signals, and the other for the autocorrelation coefficients (AC) of the stochastic codebook search target. In addition, the stochastic codebook is classified into K subcodebooks, and with these two classifiers it is possible to determine dynamically which subcodebook needs to be searched. This method achieves a reduction in the search procedure by a factor of 2-8. The efficiency of these two classifiers is discussed and the comparison of the performance between the fixed partial search and the proposed technique is also addressed.

A fast determination of stochastic excitation without codebook search in CELP coder

The major drawback of the code excitation linear prediction (CELP) coder is computational complexity that finds the best excitation vector from a stochastic codebook. To provide a synthesized speech signal with reasonable quality, the size of the stochastic codebook should be large. For this reason, the search becomes highly complex. To overcome this difficulty, several methods have been proposed. In this paper, we consider a method that enables us to directly determine the stochastic excitation vector without a codebook search. The stochastic excitation vector has been determined by projection onto a subspace that is obtained using the Karhunen-Loeve (K-L) expansion and the spectral property of the random excitation residual vector. Since the excitation vector can be determined without a codebook search, the computational complexity becomes low. From experimental results, it is shown that the proposed coder provides a synthesized speech signal that is quite comparable in quality to that of the conventional CELP coder with low computational complexity.

Speech coding and decoding system

A speech coding and decoding system, the system is operated under a known code-excited linear prediction (CELP) coding method. The CELP coding is achieved by selecting an optimum pitch vector P from an adaptive codebook and the corresponding first gain, and at the same time, selecting an optimum code vector from a stochastic codebook and the corresponding second gain. The system of the present invention is featured by a weighted orthogonalization transforming unit introduced therein. The perceptually weighted code vector AC is not used as is, as usual, but after the transformation thereof into a perceptually weighted code vector AC' by the above unit; the vector AC' being made orthogonal to the optimum perceptually weighted pitch vector AP.

Efficient codebook search for CELP vocoders

A new way of determining correlation coefficients for stochastic codebook vectors for CELP coding of speech takes advantage of the sparsely populated nature of stochastic codebook vectors. N valued input signals (e.g., convolution vectors) to be correlated with N valued codebook vectors are fed to an N by N multiplexer or other selection means and the signal values either passed to an accumulator or not according to the state of N select inputs or other identification means determined from a memory store (e.g., an EPROM) whose entries correspond to the non-zero values of the codebook vectors. The accumulator output is the correlation of the codebook vector with the input signal. A sequencer steps through the entire codebook to provide correlation values for each vectors. The results are used to determine the optimum stochastic codebook vector for replicating the particular speech frame being analyzed.

Predictive coding of speech using microphone/speaker adaptation and vector quantization

A problem with speech coders based on trained quantizers is their lack of robustness against variations in the microphone and input filter response. In this paper a simple backward adaptive prefiltering technique is proposed as a means of improving the robustness and quality of a speech coder at no cost in bit rate. The technique is particularly useful in conjunction with vector quantization (VQ) of the linear predictive coding (LPC) parameters. The performance of the prefilter, denoted a microphone and speaker adaptation (MSA) filter, has been evaluated in terms of prediction gain and spectral distortion, together with objective and subjective quality of a 7.5 kbit/s CELP speech coder. In this coder a 10-bit direct VQ of the LPC parameters using the residual energy distortion measure has been applied. This is consistent with the covariance method of LPC analysis. Simulation results illustrate that the MSA filter significantly improves the performance and robustness of the LPC VQ against changes in the input response. The 7.5 kbit/s CELP with a trained excitation codebook and MSA was found to be clearly better (subjectively and objectively) than the one without MSA. The coder with MSA also showed to be practically indistinguishable from the same CELP with unquantized LPC coefficients and a stochastic excitation codebook. >

Wideband speech coding at 16 kbit/s for a videophone application

This paper presents results on wideband 7 kHz speech coding at 16 kbit/s where the proposed CELP algorithm is implementable on a single floating point DSP. As a basic coder structure, the long-term predictor is implemented as an adaptive codebook, while a sparse Gaussian codebook with non-overlapping vectors is used for the stochastic excitation. In order to meet the complexity requirements, several methods for efficient codebook search are adopted. With these methods, it is shown that the computational effort for the basic coder structure can be reduced to 12.4 MIPS with a 7 bit stochastic codebook. A two-stage hierarchical search through the adaptive codebook is investigated. This search method reduces the computational effort further although at the cost of a small degradation in coder performance. The coder is evaluated in an absolute category rating (MOS) test using both a hi-fi handset and a loudspeaker, and compared to the CCITT standard coder G.722 at 48–64 kbit/s. The speech quality with the basic CELP structure is judged to be comparable to the G.722 coder at 48 kbit/s.

Fast methods for the CELP speech coding algorithm

Special fast procedures for the code excited linear predictive coding (CELP) algorithm have been developed to make implementation on modest hardware possible. The advantages, as well as the disadvantages, of the various fast procedures are discussed. A general formalism for the algorithm is developed, followed by the discussion of the individual procedures which are grouped according to their features. Along with the computational complexity of each procedure, its storage requirement and numerical accuracy are discussed. A large number of the fast procedures are designed to search through a particular type of codebook (most of the codebooks are stochastic in character, while a few are deterministic). Other fast procedures can be used for arbitrary codebooks and are thus also applicable to trained codebooks. Some of the fast procedures designed for stochastic codebooks can also be used for the computation of the closed pitch loop parameters, which can be interpreted as a search through a time-dependent codebook. >

An efficient stochastically excited linear predictive coding algorithm for high quality low bit rate transmission of speech

The Stochastically Excited Linear Prediction (SELP) algorithm for speech coding offers good performance at bit rates as low as 4.8 kbit/s. Linear Predictive Coding (LPC) techniques remove the short-term correlation from the speech. A pitch loop removes long-term correlation, producing a noise-like residual, which is vector quantized. Information describing the LPC filter coefficients, the long-term predictor, and the vector quantization is transmitted. In this paper, we describe improvements to the SELP algorithm which result in better speech quality and higher computational efficiency. In its closed-loop form, the pitch loop can be interpreted as a vector quantization of the desired excitation signal with an adaptive codebook populated by previous excitation sequences. To better model the non-stationarity of speech we extend this adaptive codebook with a special set of candidate vectors which are transform of other codebook entries. The second stage vector quantization is performed using a fixed stochastic codebook. In its original form, the SELP algorithm requires excessive computational effort. We employ a new recursive algorithm which performs a very fast search through the adaptive codebook. In this method, we modify the error criterion, and exploit the resulting symmetries. The same fast vector quantization procedure is applied to the stochastic codebook.

On comparing tree and codebook coding of LPC residuals at very low bit rates

It has been suggested that codebook coding of LPC residuals offers significant performance advantage over tree coding at very low rates [cf. M. R. Schroeder and B. S. Atal, Proc. ICASSP 85]. An innovation codebook provides a better coveting of the LPC innovation space (in the sense of having a smaller covering radius) than an innovation tree of the corresponding rate. These comparisons usually assume that the branches of the innovation tree are populated with independent identically distributed random variates (i.e., a stochastic tree). In this paper, the methods of selecting the innovation elements on a binary innovation tree such that the branches contain “complementary” sequences are examined. The covering properties of these “complementary” tree codes are compared with the covering properties of stochastic trees and stochastic codebooks. The results on the “complementary” tree codes indicate that (1) the performance of an innovation code did not correlate with its covering radius and (2) tree coding o...