A locally weighted distance measure for example based speech recognition

Abstract

State-of-the-art speech recognition relies on a state-dependent distance measure. In HMM systems, the distance measure is trained into state-dependent covariance matrices using a maximum likelihood or discriminative criterion. This "automatic" adjustment of the distance measure is traditionally considered an inherent advantage of HMMs over DTW recognizers, as those typically rely on a uniform Euclidean distance. In this paper we show how to incorporate a non-uniform weighted distance measure into an examplebased recognition system. By doing so we manage to combine the superior segmental behaviour of DTW with the near-optimal acoustic distance measure as found in HMMs. The non-uniform distance measure enforces modifications to the k nearest neighbours search, an essential component in our large vocabulary DTW approach. We show that the complexity of our solution remains within bounds. The validity of the full approach is verified by experimental results on the Resource Management and TIDigits tasks.

Outlier Correction for Local Distance Measures in Example Based Speech Recognition

“…In [2] we showed that a class based weighted distance measure improves significantly over the more traditional Euclidean distance metric. In this paper we refine this approach with some ideas borrowed from the domain of non-parametric density estimation that explicitly compensate for the position of a reference vector in its class.…”

“…dynamic time warping (DTW), template based recognition or episodic modeling-as an alternative to HMMs [2,3,4,5]. Example based recognition offers a natural solution to some of the problems HMMs face, especially those related to long-term sequential modeling [5].…”

Class-Dependent Dissimilarity Measures for Multiple Instance Learning

“…More attention to class-dependent distances is given in [10], where the goal is to learn weights for each feature/class combination, and to use these weights in a Mahalanobis-type metric. A similar approach is taken in [5] to improve performance in speech recognition. In [19], the authors propose learning different metrics for different classes and show that this improves classification results.…”

“…This may not be completely natural for a MIL problem, because we have some information about how positive bags are different from negative bags. In supervised learning problems where classes are expected to behave differently, class-dependent distances [10,5,19] or features [2,8] have been suggested. In this work we examine whether a similar approach might be reasonable for MIL problems.…”

Posterior-Based Features and Distances in Template Matching for Speech Recognition

“…In the TM-based approach, investigation has been recently carried out to estimate the parameters of the weighting matrix of the Mahalanobis distance to improve the performance. A maximum-likelihood estimation was described in [12] and a discriminative procedure was presented in [13]. However, these methods require a large amount of data to properly estimate the weights.…”