Robust ASR using Support Vector Machines

Abstract

The improved theoretical properties of Support Vector Machines with respect to other machine learning alternatives due to their max-margin training paradigm have led us to suggest them as a good technique for robust speech recognition. However, important shortcomings have had to be circumvented, the most important being the normalisation of the time duration of different realisations of the acoustic speech units.In this paper, we have compared two approaches in noisy environments: first, a hybrid HMM-SVM solution where a fixed number of frames is selected by means of an HMM segmentation and second, a normalisation kernel called Dynamic Time Alignment Kernel (DTAK) first introduced in [1] and based on DTW (Dynamic Time Warping). Special attention has been paid to the adaptation of both alternatives to noisy environments, comparing two types of parameterisations and performing suitable feature normalisation operations. The results show that the DTA Kernel provides important advantages over the baseline HMM system in medium to bad noise conditions, also outperforming the results of the hybrid system.

SVMs for Automatic Speech Recognition: A Survey

“…In summary, the reported results [31] show that SVMs exhibit a robust behaviour, as expected. In particular, the DTAK-based system turns out to be effective in noisy scenarios.…”

“…Results for the well-known SpeechDat-4000 database are presented in [31]. The whole database is not used: specifically, only the isolated-digit utterances are used for the experiments.…”

“…In the remaining conditions, which correspond to medium SNRs, the system performances do not exhibit statistically significant differences. Please refer to [31] for more details about the DTAK-based system and the experimental setup.…”

“…Another possible solution is showed in [28,30,31], where the non-uniform distribution of analysis instants provided by the internal states transitions of an HMM with a fixed number of states and a Viterbi decoder is used for dimensional normalization. The rationale behind this proposal is that the uniform resampling methods are produced without any consideration about the information (or lack of information) that speech analysis segments were providing.…”

Automatic fingersign-to-speech translation system

“…Recent automatic speech recognizers exploit mathematical techniques such as Hidden Markov Models (HMM) [61], Artificial Neural Networks (ANN) [51], Dynamic Bayesian Networks (DBN) [53], Dynamic Time Warping (DTW) or dynamic programming [24], Support Vector Machines (SVM) [52] or some hybrid models [18,55]. The most popular ASR models apply speaker independent speech recognition, though in some cases (for instance, personalized systems that have to recognize owner only) speaker dependent systems are more adequate, as in the case of personalized systems.…”