On the effects of varying filter bank parameters on isolated word recognition

Abstract

The vast majority of commercially available isolated word recognizers use a filter bank analysis as the front end processing for recognition. It is not well understood how the parameters of different filter banks (e.g., number of filters, types of filters, filter spacing, etc.) affect recognizer performance. In this talk we present results of performance evaluation of several types of filter bank analyzer in a speaker trained, isolated word recognition test using dialed-up telephone line recordings. We have studied both DFT (discrete Fourier transform) and direct form implementations of the filter banks. We have also considered uniform and nonuniform filter spacings. The results indicate that the best performance (highest word accuracy) is obtained by both a 15-channel uniform filter bank and a 13-channel nonuniform filter bank (with channels spaced along a critical band scale). In comparison to a conventional linear predictive coding (LPC) word recognizer, the performance of the best filter bank recognizers was, on average, several percent worse than that of an eighth order LPC-based recognizer.

A comparative study of mel cepstra and EIH for phone classification under adverse conditions

“…While the energy based spectral analysis schemes such as LPC work well under similar acoustic training and testing conditions, the system performance deteriorates significantly under adverse signal conditions [4]. For example, for an alphanumeric recognition task, the performance of the SPHINX system falls from 77-85% accuracy with matched training and testing recording environments to 19-37% accuracy on cross conditions [1].…”

Hidden Markov Models for Speech Recognition

“…These include such standard methods as measurement of the discrete (fast) Fourier transform (FFT), all-pole minimum-phase linear prediction (LPC) methods, and autoregressive/moving average models (Allen and Rabiner 1977;Atal and Hanauer 1971;Cadzow 1982;Makhoul 1975;Markel and Gray 1976;Schafer and Rabiner 1971). Even the more traditional filter-bank method of spectral analysis is still used in some systems (Dautrich, Rabiner, and Martin 1983), particularly in hardware implementations. To emphasize spectral properties that are known to be important to a human listener, auditory models can be incorporated in the overall spectral representation (Cohen 1985;Ghitza 1986).…”

Hidden Markov model training with contaminated speech material for distant-talking speech recognition

“…With suitable generation of noisy training data, an upper bound in system performance can be obtained with respect to many feature/model based compensation techniques (Dautrich et al, 1983;Gales, 1995;Gong, 1995). In particular, training data contamination may be conveniently adopted where the speech distortion in the real environment is unknown or would be difficult to model explicitly (Acero, Deng, Kristjansson & Zhang, 2000).…”

“…A different approach, known in the literature as training data contamination (Dautrich, Rabiner & Martin, 1983;Das, Bakis, Nadas, Nahamoo & Picheny, 1993;Gong, 1995), provides a way of training acoustic models which are more robust and representative of the given real noisy environment than those derived by training on the corresponding clean speech. In general, training of speech recognizers is accomplished by using large speech corpora.…”