Investigations into the Crandem Approach to Word Recognition

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 725–728, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Investigations into the Crandem Approach to Word Recognition Rohit Prabhavalkar, Preethi Jyothi, William Hartmann, Jeremy Morris, and Eric Fosler-Lussier Department of Computer Science and Engineering The Ohio State University, Columbus, OH {prabhava,jyothi,hartmanw,morrijer,fosler}@cse.ohio-state.edu Abstract We suggest improvements to a previously pro- posed framework for integrating Conditional Random Fields and Hidden Markov Models, dubbed a Crandem system (2009). The pre- vious authors’ work suggested that local la- bel posteriors derived from the CRF were too low-entropy for use in word-level automatic speech recognition. As an alternative to the log posterior representation used in their sys- tem, we explore frame-level representations derived from the CRF feature functions. We also describe a weight normalization transfor- mation that leads to increased entropy of the CRF posteriors. We report significant gains over the previous Crandem system on the Wall Street Journal word recognition task. 1 Introduction Conditional Random Fields (CRFs) (Lafferty et al., 2001) have recently emerged as a promising new paradigm in the domain of Automatic Speech Recognition (ASR). Unlike Hidden Markov Mod- els (HMMs), CRFs are direct discriminative models: they predict the probability of a label sequence con- ditioned on the input. As a result, CRFs can capture long-range dependencies in the data and avoid the need for restrictive independence assumptions. Vari- ants of CRFs have been successfully used in phone recognition tasks (Gunawardana et al., 2005; Morris and Fosler-Lussier, 2008; Hifny and Renals, 2009). While the improvements in the phone recognition task are encouraging, recent efforts have been di- rected towards extending the CRF paradigm to the word recognition level (Zweig and Nguyen, 2009; Morris and Fosler-Lussier, 2009). The Crandem system (Morris and Fosler-Lussier, 2009) is one of the promising approaches in this regard. The Cran- dem system is directly inspired by the techniques of the Tandem system (Hermansky et al., 2000), where phone-label posterior estimates produced by a Multi-Layer Perceptron (MLP) are transformed into a suitable acoustic representation for a standard HMM. In both systems, the frame-based log poste- rior vector of P(phone|acoustics) over all phones is decorrelated using the Karhunen-Loeve (KL) trans- form; unlike MLPs, CRFs take into account the en- tire label sequence when computing local posteriors. However, posterior estimates from the CRF tend to be overconfident compared to MLP posteriors (Mor- ris and Fosler-Lussier, 2009). In this paper, we analyze the interplay between the various steps involved in the Crandem process. Is the local posterior representation from the CRF the best representation? Given that the CRF poste- rior estimates can be overconfident, what transfor- mations to the posteriors are appropriate? In Section 2 we briefly describe CRFs and the Crandem framework. We suggest techniques for im- proving Crandem word recognition performance in Section 3. Details of experiments and our results are discussed in Sections 4 and 5 respectively. We con- clude with a discussion of future work in Section 6. 2 CRFs …

Content truncated to 500 words