. In this paper, we discuss a new automatic speech recognition (ASR) approach based on independent processing and recombination of several feature streams. In this framework, it is assumed that the speech signal is represented in terms of multiple input streams, each input stream representing a different characteristic of the signal. If the streams are entirely synchronous, they may be accommodated simply (as they usually are in state-of-the-art systems). However, as discussed in the paper, it may be required to permit some degree of asynchrony between streams. This paper introduces the basic framework of a statistical structure that can accommodate multiple (asynchronous) observation streams (possibly exhibiting different frame rates). This approach will then be applied to the particular case of multi-band speech recognition and will be shown to yield significantly better noise robustness. 2 IDIAP--RR 96-07 1 Introduction In current automatic speech recognition (ASR) systems, the a...

Abstract not found

The hybrid Hidden Markov Model (HMM) / Neural Network (NN) speech recognition system at the International Computer Science Institute (ICSI) uses a single hidden layer MLP (Multi Layer Perceptron) to compute the emission probabilities of the states of the HMM. This recognition approach was developed and has traditionally been used for large vocabulary size continuous speech recognition. In this report, however, such a recognition scheme is applied directly to three much smaller vocabulary size corpora, the Bellcore isolated digits, the TI connected digits, and the Center for Spoken Language Understanding Numbers'93 database. The work reported here is not only on developing small baseline systems to facilitate all future research experiments, but also on using these systems to evaluate front-end research issues, and the feasibility of using context-dependency for speech recognition under the hybrid approach developed at ICSI. In addition, using the TI connected digits, the performance of...

Recently it has been shown that Artificial Neural Networks (ANNs) can be used to augment speech recognizers whose underlying structure is essentially that of Hidden Markov Models (HMMs). In particular, we have shown that fairly simple layered structures, which we lately have termed Big Dumb Neural Networks (BDNNs), can be discriminatively trained to estimate emission probabilities for HMMs. Many (relatively simple) speech recognition systems based on this approach, and generally referred to as hybrid HMM/ANN systems, have been proved, on controlled tests, to be both e ective in terms of accuracy and e cient in terms of CPU and memory run-time requirements. In this paper, we discuss some current research topics on extending these results to somewhat more complex systems, including new theoretical and experimental developments on transition-based recognition systems and training of HMM/ANN hybrids to maximize the global posterior probabilities.

Abstract not found