The authors describe typical variabilities that are found in telephone channels and show how these variabilities impact automatic speech recognition (ASR)-system performance. They describe an algorithm, based on spectral subtraction, that compensates for these types of degradations. This algorithm can be utilized at the beginning of ASR interactions to estimate a compensation vector that improves recognition performance during the rest of the interaction. They have tested and validated the algorithm in terms of recognition performance and speed of convergence.>

Describes a speech coding algorithm termed stochastic tree coding (STC) that is capable of delivering high-quality speech at bit rates of 4.8 to 8 kb/s. Through the use of a tree-structured excitation codebook and an efficient 3-tap pitch search, the algorithm has a moderate computational complexity that enables implementation on a fixed-point DSP56001. The authors also describe an error protection scheme designed for wireless fading channels.>

Algorithms for use in the estimation of left-right hidden Markov models with skip transitions are described. The results of several large-scale experiments conducted to illustrate the effects of some problems related to the use of the algorithms are presented. Solutions are presented to resolve the problems. The results show that the improvements result in more accurate modeling of training data.>

Methods of increasing the performance of two components of stochastic coding, the codebook excitation search and the long-term synthesis filter, while maintaining or reducing complexity are discussed. A flexible tree-coding technique is introduced that significantly lowers the complexity of the codebook search with very little loss in quality. An efficient method of performing a three-tap long-term prediction analysis in a closed loop is developed. It improves perceptual quality and requires little additional complexity. The result is a stochastic coder which produces high-quality speech at 8 kb/s with manageable computational complexity.>

The authors report the development of an error control scheme for the adaptive subbands excited transform coding algorithm. It is a relatively simple source-channel coding scheme which takes into account the sensitivity of the compression parameters to transmission errors. The added computational complexity, encoding-decoding delay, and rate overhead are all relatively small. The compression algorithm, with error control and without, has been implemented and tested at a total bit rate of 16 kb/s. The channel used is the memoryless binary symmetric channel with error probability up to 0.01. The results show that under error-free conditions speech quality is unchanged. However, under error conditions the quality of speech produced by the protected algorithm is consistently better.>

This paper describes a very effective method for reducing framing noise associated with block-by-block speech coding algorithms. It is a post-coding speech enhancement procedure designed specifically to combat this type of noise. The procedure is based on a fully adaptive comb filter which is adapted in synchrony with the coding algorithm's frame boundary. An optimal (minimum mean-squared-error sense) linear prediction approach is applied to determine the filter coefficients and adapt them to the short-time statistics of the speech. A formal subjective listening trial has shown the method to significantly improve the perceived quality of block-coded speech.

This paper describes a medium-bit-rate speech compression algorithm for telephone applications. The basic configuration of the adaptive subbands excited transform (ASET) coder can be described as frequency domain residual subbands coding, a hybrid of transform coding and residual excited coding. The algorithm outperforms traditional RELP (residual excited linear predictive) coders for both speech and non-speech signals. The algorithm is much less complex than adaptive transform coding (ATC) algorithms, yet at 16-kb/s it produces comparable speech quality. A multichannel hardware implementation of the algorithm has been developed and is reported in [1].

This paper describes the development of a new approach to hardware implementation of speech compression algorithms. In considering the hardware architecture much consideration was given to future integration into the public telephone system or other multiuser systems. We propose a multichannel multiprocessor system placed at the central office rather than the traditionally considered single channel DSP-chip system. Dedicated processors and distributed and parallel processing are basic elements in the system architecture. The proposed approach was used to develop an experimental real-time processor for 16-kb/s compression, implementing the adaptive subbands excited transform (ASET) coding algorithm [1]. A system capable of serving up to 24 channels has been successfully built and demonstrated. In addition to high-quality speech, the experimental system results in a cost per channel end that is approximately one-third the cost of a DSP-chip implementation.

For a class of stationary Gaussian sources and the squared-error distortion measure, an asymptotically optimal tree coding scheme is derived using tree codes with finite branch length. The distribution of the reproduction process is derived in an explicit form, and using the random coding argument and results from the theory of branching processes with stationary ergodic environmental processes, the source coding theorem is proved. The theorem is applicable to all Gaussian sources with a power spectrum that satisfies the Lip 1 condition with no restriction on the coding rate.

This paper presents a rate-distortion derived transform trellis coding (TTC) scheme with applications to Gaussian AR sources and speech data. The optimal encoder consists of a Karhunen-Loeve transform (KLT) on the source output, followed by a search on a trellis structured random code, where the decoder is a time-variant nonlinear filter. The scheme is implementable and applicable to stationary Gaussian sources with a bounded and continuous power spectrum and the squared error distortion measure. The code construction is based on the power or eigenvalue spectrum of the source with no restriction on the coding rate. The TTC scheme is first applied to encode a Gaussian AR source often used to model speech. Simulations were conducted at several rates, using an optimal KLT and the suboptimal discrete cosine transform (DCT). Results demonstrate that the DCT performs as well as the KLT, and both yield average distortions very close to the distortion-rate function. For speech data, an adaptive version of the DCT TTC scheme is applied to encode two speech sentences at several coding rates. The adaptation is controlled by an estimate of the short-term eigenvalue spectrum which is transmitted as side information to the receiver. The proposed scheme is a very efficient speech waveform coder that provides reconstructed speech with very high signal-to-noise ratio values and very good perceptual quality at low bit rates.

For stationary discrete-time Gaussian sources and the squared-error distortion measure, a trellis source code is constructed. The encoder consists of a Karhunen-Loeve transform on the source output followed by a search on a trellis structured code, where the decoder is a time-variant nonlinear filter. The corresponding code theorem is proved using the random coding argument. The proof technique follows that of Viterbi and Omura, who proved the trellis coding theorem for memoryless sources. The resultant coding scheme is implementable and applicable at any nonzero rate to a stationary Gaussian source with a bounded and continuous power spectrum. Therefore. for stationary sources, it is more general than Berger's tree coding scheme, which is restricted to autoregressive Gaussian sources in a region of high rate (low distortion).