Learning Structured Models for Phone Recognition

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 897–905, Prague, June 2007. c⃝2007 Association for Computational Linguistics Learning Structured Models for Phone Recognition Slav Petrov Adam Pauls Dan Klein Computer Science Department, EECS Divison University of California at Berkeley Berkeley, CA, 94720, USA {petrov,adpauls,klein}@cs.berkeley.edu Abstract We present a maximally streamlined approach to learning HMM-based acoustic models for automatic speech recognition. In our approach, an initial mono- phone HMM is iteratively refined using a split-merge EM procedure which makes no assumptions about subphone structure or context-dependent structure, and which uses only a single Gaussian per HMM state. Despite the much simplified training process, our acoustic model achieves state-of-the-art results on phone classification (where it outperforms almost all other methods) and competitive performance on phone recognition (where it outperforms standard CD triphone / subphone / GMM approaches). We also present an analysis of what is and is not learned by our system. 1 Introduction Continuous density hidden Markov models (HMMs) underlie most automatic speech recognition (ASR) systems in some form. While the basic algorithms for HMM learning and inference are quite general, acoustic models of speech standardly employ rich speech-specific structures to improve performance. For example, it is well known that a monophone HMM with one state per phone is too coarse an approximation to the true articulatory and acoustic process. The HMM state space is therefore refined in several ways. To model phone-internal dynam- ics, phones are split into beginning, middle, and end subphones (Jelinek, 1976). To model cross-phone coarticulation, the states of the HMM are refined by splitting the phones into context-dependent tri- phones. These states are then re-clustered (Odell, 1995) and the parameters of their observation dis- tributions are tied back together (Young and Wood- land, 1994). Finally, to model complex emission densities, states emit mixtures of multivariate Gaus- sians. This standard structure is shown schemati- cally in Figure 1. While this rich structure is pho- netically well-motivated and empirically success- ful, so much structural bias may be unnecessary, or even harmful. For example in the domain of syn- tactic parsing with probabilistic context-free gram- mars (PCFGs), a surprising recent result is that au- tomatically induced grammar refinements can out- perform sophisticated methods which exploit sub- stantial manually articulated structure (Petrov et al., 2006). In this paper, we consider a much more automatic, data-driven approach to learning HMM structure for acoustic modeling, analagous to the approach taken by Petrov et al. (2006) for learning PCFGs. We start with a minimal monophone HMM in which there is a single state for each (context-independent) phone. Moreover, the emission model for each state is a sin- gle multivariate Gaussian (over the standard MFCC acoustic features). We then iteratively refine this minimal HMM through state splitting, adding com- plexity as needed. States in the refined HMMs are always substates of the original HMM and are there- fore each identified with a unique base phone. States are split, estimated, and (perhaps) merged, based …

Content truncated to 500 words