Template-Based Continuous Speech Recognition

Abstract

Despite their known weaknesses, hidden Markov models (HMMs) have been the dominant technique for acoustic modeling in speech recognition for over two decades. Still, the advances in the HMM framework have not solved its key problems: it discards information about time dependencies and is prone to overgeneralization. In this paper, we attempt to overcome these problems by relying on straightforward template matching. The basis for the recognizer is the well-known DTW algorithm. However, classical DTW continuous speech recognition results in an explosion of the search space. The traditional top-down search is therefore complemented with a data-driven selection of candidates for DTW alignment. We also extend the DTW framework with a flexible subword unit mechanism and a class sensitive distance measure-two components suggested by state-of-the-art HMM systems. The added flexibility of the unit selection in the template-based framework leads to new approaches to speaker and environment adaptation. The template matching system reaches a performance somewhat worse than the best published HMM results for the Resource Management benchmark, but thanks to complementarity of errors between the HMM and DTW systems, the combination of both leads to a decrease in word error rate with 17% compared to the HMM results.

Noise robust exemplar matching with alpha–beta divergence

“…(De Wachter et al, 2007)) in the scope of this paper, which would immediately enhance the recognition performance at higher SNR levels. In such a setting, the acoustic scores obtained from both streams can be combined to benefit from the noise robustness of exemplar-based acoustic modeling and better discrimination of the statistical models such as complex GMM distributions in conjunction with MFCC features or DNNs.…”

“…Considering the dimensionality and computational restrictions, the same framework using exemplars associated with more general subword units such as phones or syllables could be applied to a medium or large vocabulary task. Only the current decoding scheme would need to be redesigned in a way that it will incorporate a language model combined with the acoustic costs, but for this it could largely rely on existing exemplar matching frameworks (De Wachter et al, 2007).…”

“…Data-driven automatic speech recognition (ASR) techniques (De Wachter et al, 2003;Aradilla et al, 2005;Deselaers et al, 2007;Sundaram and Bellegarda, 2012;Sainath et al, 2012;Heigold et al, 2012;Sun et al, 2014) became popular in the last decade as a viable alternative after the long dominance of statistical acoustic modeling in the form of the Gaussian mixture models (GMM) in hidden Markov models (HMM) (Bourlard et al, 1996). Templates or exemplars are labeled speech segments of multiple lengths extracted from training data, each associated with a certain class, i.e.…”

“…Exemplar matching-based recognition can be performed by evaluating the similarity of the exemplars with the segments from the input speech with respect to a distance/divergence metric by applying dynamic time warping (Sakoe and Chiba, 1971;Ney and Ortmanns, 1999;De Wachter et al, 2007). In these applications, speech is represented using discriminatively trained features to ensure that the used distance/divergence metric mostly yields lower scores for the matching class compared to the other classes, resulting in increased recognition accuracies.…”

A tighter lower bound estimate for dynamic time warping

“…In the history of automatic speech recognition (ASR), DTW first became popular in isolated and connected word recognition and then was supplanted by hidden Markov models (HMMs), a statistical modeling framework appropriate for large vocabulary continuous speech recognition (LVCSR). However, DTW has drawn much interest recently for unsupervised and low-resource tasks, e.g., template-based speech recognition [2,3], unsupervised speech pattern discovery [4,5], example-based spoken term detection (STD) [6,7] and acousticbased spoken document segmentation [8]. Recently, Zhang et.…”