Marcel Vasilache

A combination of discriminative and maximum likelihood techniques for noise robust speech recognition

We study how discriminative and maximum likelihood (ML) techniques should be combined in order to maximize the recognition accuracy of a speaker-independent automatic speech recognition (ASR) system that includes speaker adaptation. We compare two training approaches for the speaker-independent case and examine how well they perform together with four different speaker adaptation schemes. In a noise robust connected digit recognition task we show that the minimum classification error (MCE) training approach for speaker-independent modelling together with the Bayesian speaker adaptation scheme provide the highest classification accuracy over the whole lifespan of an ASR system. With the MCE training we are capable of reducing the recognition errors by 30% over the ML approach in the speaker-independent case. With the Bayesian speaker adaptation scheme we can further reduce the error rates by 62% using only as few as five adaptation utterances.

On a practical design of a low complexity speech recognition engine

We outline the main design features of a low complexity speech recognition engine targeted for mobile devices. Although major parts have already been presented, new features and important refinements of the original ideas, which were omitted, are now described. We also show how these techniques can be successfully combined in order to achieve various design targets with minimized impact on the recognition performance.

Predictive multiple-scale lattice VQ for LSF quantization

This paper introduces a new lattice quantization scheme, the multiple-scale lattice vector quantization (MSLVQ), based on the truncation of the D/sub 10//sup +/ lattice. The codebook is composed of several copies of the truncated lattice scaled with different scaling factors. A fast nearest neighbor search is introduced. We compare the performance of predictive MSLVQ for quantization of line spectrum frequency (LSF) coefficients with the quantization technique used in the G.729 codec and show the better performance of our method in terms of spectral distortion. The MSLVQ scheme achieves the transparent quality at 21 bits/frame.

Multi-rate hmm quantization for speech recognition

This paper refines the idea of scalar quantization for hidden Markov model (HMM) parameters which was introduced in an earlier contribution. With the proposed multi-rate approach it is shown that an increased model compression can be achieved with a significant computational complexity reduction while also closely preserving the recognition performance of the original models.