Hmm Mixtures Hmm2 For Robust Speech Recognition

Author : Marta Casar
Publisher :
ISBN 13 : 9783902613080
Total Pages : pages
Book Rating : 4.6/5 (13 download)

DOWNLOAD NOW!

Download or read book Double Layer Architectures for Automatic Speech Recognition Using HMM. written by Marta Casar and published by . This book was released on 2007 with total page pages. Available in PDF, EPUB and Kindle. Book excerpt: The future of speech-related technologies is connected to the improvement of speech recognition quality. Until recently, speech recognition technologies and applications had assumed that there were certain limitations regarding vocabulary length, speaker independence, and environmental noise or acoustic events. In the future, however, ASR must deal with these restrictions and it must also be able to introduce other speech-related non-acoustic information that is available in speech signals. Furthermore, HMM-based statistical modelling--the standard state-of-the-art ASR--has several time-domain limitations that are known to affect recognition performance. Context is usually represented by means of spectral dynamic features (namely, its first and second derivatives). However, they are generally limited to a few subsequent feature vectors and do not represent long-term variations. To overcome all these drawbacks and to achieve a qualitative improvement in speech recognition, a change of paradigm from conventional speech recognizers has been proposed by several authors. Although some authors propose a move away from HMM-based recognition (or, at the very least, introducing hybrid solutions), we are adhering to Markovbased acoustic modelling as we believe its approach is still unbeatable. However, to overcome HMM-related limitations certain innovative solutions are required. Throughout this chapter we have pointed out different approaches for improving standard HMM-based ASR systems. The main solutions for modelling temporal evolution and speech trajectory have been introduced, together with some ideas on how second-order HMMs deal with the same problems. These models provide an improvement in most cases, but they also require major modifications in the decoding of algorithms. Generally, there is also a considerable increase in complexity, even if this is compensated for by a moderate gain. Layered architectures have been presented, and special attention has been paid to the implementation of the second layer using extended HMMs. Two implementations for this second layer have been described in detail. The first relies on modelling the temporal evolution of acoustic HMM state scores. In the second one, the evolution of the acoustic HMM is modelled by the speech utterance as a new way of modelling state transitions. This can be done in two ways, namely, by taking into account or ignoring the context of each active state while the "path" followed by the speech signal through the HMM states is being modelled. Again, speech recognition performance improves that of a conventional HMMbased speech recognition system, but at the cost of increased complexity. Although current research solutions should not be unduly concerned by the computational cost (due to the constant increase in the processing capacity of computers), it is important to keep their implementation in commercial applications in mind. Therefore, a great deal of work remains if layered architectures are to be generalized for large vocabulary applications that keep complexity down to a moderate level. Efforts should be made in the field of research for defining and testing innovative approaches to implementing layered architectures. Although keeping an HMM-based scheme for the different layers reduces the overall complexity, a change in paradigm may help to bring about significant improvements.