Yasuji Ohta

Speech coding and decoding system

A CELP type speech coding system is provided with an arithmetic processing unit which transforms a perceptual weighted input speech signal vector AX to a vector t AAX, a sparse adaptive codebook which stores a plurality of pitch prediction residual vectors P sparsed by a sparse unit, and a multiplying unit which multiplies the successively read out vectors P and the output t AAX from the arithmetic processing unit. In addition, the CELP type speech coding system includes a filter operation unit which performs a filter operation on the vectors P, and an evaluation unit which finds the optimum vector P based on the output from the filter operation unit, so as to enable reduction of the amount of arithmetic operations.

Speech encoding/decoding apparatus having selected encoders

Several encoders perform a local decoding of a speech signal and extract excitation information and vocal tract information from a speech signal for an encoding operation. The transmission rate ratio between the excitation information and the vocal tract information are different for each encoder. An evaluation/selection unit evaluates the quality of decoded signals subjected to a local decoding in each of the encoders, determines the most suitable encoders from among the several encoders based on the result of the evaluation, and selects the most suitable encoder, thereby outputting the selection result as selection information. The decoder decodes a speech signal based on selection information, vocal tract information and excitation information. The evaluation/selection unit selects the output from the encoder in which the quality of a locally decoded signal is the most preferable. When vocal tract information changes little, the vocal tract information is not output, thereby allowing for increased quality of information. As much of the surplus of unused vocal tract information as possible is assigned to a residual signal. Thus, the quality of a decoded speech signal is improved.

Pitch sharpening for perceptually improved CELP, and the sparse-delta codebook for reduced computation

The authors introduce two techniques for improving low-bit-rate CELP (code excited linear prediction) speech coders. The sparse-delta stochastic codebook is a recursive codebook design which can be searched using roughly 5% of the computational load required to search a full Gaussian codebook. Pitch sharpening is a class of algorithms which attempt to improve the perceptual quality of CELP by limiting the feedback of low-amplitude noiselike information to the adaptive codebook. Simulation results are presented for sparse-delta, ternary sparse-delta, and simplified-search sparse-delta coders, and for reduced-gain and sparse-adaptive-codebook pitch sharpening algorithms.<<ETX>>

Speech coding system having codebook storing differential vectors between each two adjoining code vectors

A speech coding system is provided where input speech is coded by finding via an evaluation computation a code vector giving a minimum error between reproduced signals obtained by linear prediction analysis filter processing, simulating speech path characteristics, on code vectors successively read out from a noise codebook storing a plurality of noise trains as code vectors and an input speech signal and by using a code specifying the code vector. In the speech coding system, the noise codebook includes a delta vector codebook which stores an initial vector and a plurality of delta vectors having difference vectors between adjoining code vectors. In addition, provision is made in the computing unit for the evaluation computation of a cyclic adding unit for cumulatively adding the delta vectors to virtually reproduce the code vectors.

Principal axis extracting vector excitation coding: high quality speech at 8 kb/s

Principal axis extracting vector excitation coding (PAVXC) is proposed in which an efficient codebook search method and an improved excitation modeling technique are used. A two-step codebook search using a small-size codebook of 20 dimensions (half the conventional size) can reduce the computational complexity by a factor of 64, maintaining the reproduced speech quality within a slight degradation of 0.9 dB in SNR seg. Multiple vector excitation using an impulse vector jointly with the conventional stochastic code vector can provide more precise representation of the excitation. The impulse vector is extracted from each code vector by principal axis extraction, and its gain is jointly optimized with the gain for the code vector. The improvement in SNR seg is 1.2 dB, which compensates for the degradation caused by the two-step codebook search. The quality of the reproduced speech using PAVXC is subjectively much better than that of conventional VXC, thanks especially to the improvement in voiced regions.<<ETX>>

Tree-structured delta codebook for an efficient implementation of CELP

The authors present a structured codebook for reducing the complexity and memory of a CELP stochastic codebook search. The tree-structured delta codebook, whose code vectors are generated from a small number of delta vectors, can be searched efficiently by calculating the vector correlations recursively. The complexity was reduced to 1/70 that of a conventional full-Gaussian codebook, and the memory for codebook storage was reduced to 1/100. Another excellent feature of this codebook is the feasibility of changing the distribution of code vectors adaptively. The codebook adaptation method (delta vector sorting) provides an SNRseg improvement of 0.6 dB, and consistent improvement in perceptual quality, by changing the order of delta vectors to fit the input speech.<<ETX>>

Structured Stochastic Codebook and Codebook Adaptation for CELP

Since its introduction in 1984, Code Excited Linear Prediction (CELP) [1] has been intensively investigated as a promising coding algorithm for providing good quality speech at low bit rates. CELP is the name for a class of coding algorithms that employs vector quantization (VQ) using a perceptually weighted error criterion measured in an Analysis-by-Synthesis loop. This process gives an efficient representation of the excitation signal and exhibits better performance than conventional coding methods. However, the codebook search requires a huge computational load, which is a major drawback in the practical implementation of CELP. In particular, for digital cellular communications, which is considered the biggest application for low bit-rate speech coding, reducing the complexity of CELP is important for small hardware size and low power consumption.

Efficient PSI-CELP DSP implementation

Pitch Synchronous Innovation-Code Excited Linear Prediction (PSI-CELP) algorithm has been adopted as a standard speech codec in the new Japanese Personal Digital Cellular (PDC) system. In this paper, we present an efficient implementation scheme for the PSI-CELP algorithm with a Digital Signal Processor (DSP). A PSI-CELP algorithm features the PSI technique, which can improve speech quality to repeat a part of a stochastic codeword with pitch period. We propose a specialized function that can support DSP instructions to allow efficient PSI implementation. According to our practical estimation with the PSI-CELP standard specification, our method can greatly reduce the costs for both parts complexity and required memory space.

New description language for DSP firmware conversion

DSPs (Digital Signal Processors) are widely used, especially in digital cellular phones. However, it is very difficult and time consuming to create firmware for DSPs. We have considered an improved firmware development system. This paper introduce a new DSP firmware description language. We demonstrate firmware conversion between two different DSPs using our new language.