The Application of Hidden Markov Models in Speech Recognition

Abstract

Hidden Markov Models (HMMs) provide a simple and effective framework for modelling time-varying spectral vector sequences. As a consequence, almost all present day large vocabulary continuous speech recognition (LVCSR) systems are based on HMMs.Whereas the basic principles underlying HMM-based LVCSR are rather straightforward, the approximations and simplifying assumptions involved in a direct implementation of these principles would result in a system which has poor accuracy and unacceptable sensitivity to changes in operating environment. Thus, the practical application of HMMs in modern systems involves considerable sophistication.The aim of this review is first to present the core architecture of a HMM-based LVCSR system and then describe the various refinements which are needed to achieve state-of-the-art performance. These refinements include feature projection, improved covariance modelling, discriminative parameter estimation, adaptation and normalisation, noise compensation and multi-pass system combination. The review concludes with a case study of LVCSR for Broadcast News and Conversation transcription in order to illustrate the techniques described.

LTSD and GDMD features for Telephone Speech Endpoint Detection

“…The phrase modeling in this application is done by a whole-phrase continuous HMM [2,10,27]. The selected model is with a left-to-right topology with no skip state and the output distributions are represented as mixture of Gaussians with diagonal covariance matrices.…”

“…The selected model is with a left-to-right topology with no skip state and the output distributions are represented as mixture of Gaussians with diagonal covariance matrices. The HMM training is carried out by well-known Baum-Welch Algorithm [10,27]. In the verification are used the individual speaker's thresholds.…”

“…In the second experiment the performance of the endpoints detection algorithms in terms of the recognition rate is estimated via two fixed-text speaker verification applications. The first application is based on the Dynamic Time Warping (DTW) algorithm [21] while the second one uses the left-to-right HMM paradigm [10]. The verification results are compared to these obtained by the manual endpoint detection.…”

Subword Modeling for Automatic Speech Recognition: Past, Present, and Emerging Approaches

“…Each context-dependent unit is typically represented by a hidden Markov model (HMM) with Gaussian mixture observation densities, which account for the remaining acoustic variation among different instances of the same unit. For further details about the architecture of standard HMM-based recognizers, see [4]. 1 Linguists distinguish phones-acoustic realizations of speech sounds-from phonemes-abstract sound units, each possibly corresponding to multiple phones, such that a change in a single phoneme can change a word's identity.…”

“…Typical sub-phonetic features are articulatory features, which may be binary or multivalued and characterize in some way the configuration of the vocal tract. 4 Roughly 80% of phonetic substitutions of consonants in the Switchboard Transcription Project data consist of a single articulatory feature change [10]. In addition, effects such as nasalization, rounding, and stop consonant epenthesis can be the result of asynchrony between articulatory trajectories [50].…”