H . Hermansky : " AUDITORY MODELING IN AUTOMATIC RECOGNITION OF SPEECH

The paper argues against blind implementing of scattered accidental knowledge which may be irrelevant to a speech recognition task and advance the notion that the reason for applying knowledge of human auditory perception in engineering applications should be the ability of perception to suppress some parts of information in the speech message. In general, it advocates selective use of auditory knowledge, optimized on real speech data. 1 Ignorance and Knowledge in Handling The Nonlinguistic Variability in ASR With an advent of powerful stochastic classiication techniques , many believe that the task of alleviating the irrelevant variability should be left on the classiier. In the training stage the recognizer would be presented with all the information that is available (i.e., both \signal" and \noise"). Such exhaustive training should allow (during recognition) for a separation of the desired signal from the noise. However, for this to be true, the classiier would have to be a model-free classiier (so that no prior knowledge constrains its capabilities) and it would have to be presented with all possible past and future speech data. Therefore, weakly-structured data-driven ASR systems require large amounts of training data. Further, the knowledge acquired by the training is typically speciic to a given application and may not be easily reusable in another system. Some speculate that more structured designs with some build-in knowledge of human speech communication processes could provide desirable constraints for the ASR problem. As will become more obvious after reading through this paper, we are currently advocating selective use of auditory knowledge optimized on real speech data. 2 Analysis in ASR The information rate in speech signal is by some estimates 1] of the order of 36 kbits/s. The written equivalent of the linguistic message in the signal is less that 50 bits/s 1]. The general task of ASR is to identify and reduce information in the signal. The analysis should support this goal by alleviating as much as possible non-linguistic factors in the signal. Additional practical requirements for current ASR techniques are that speech features derived from speech analysis should be low-dimensional and their statistics should be well described by gaussian distributions. Automatic speech recognition (ASR) typically uses features based on a short-term spectrum of speech which describes time-varying speech signal as a sequence of short-term feature vectors. Each vector reeects properties of a relatively short (10-20 ms) segment of the signal. Each individual feature vector is …