Experiments with Tree-Structured MMI Encoders on the RM Task

Experiments with Tree-Structured MMI Encoders on the RM Task Mark T. Anikst, William S. Meisel, Matthew C. Stares Speech Systems Incorporated 18356 Oxnard Street Tarzana, California 91356 Kai-Fu Lee Carnegie Mellon University Pittsburgh, Pennsylvania 15213 ABSTRACT This paper describes the tree-structured maximum mutual information (MMI) encoders used in SSrs Phonetic Engine ® to perform large-vocabulary, continuous speech recognition. The MMI encoders are arranged into a two-stage cascade. At each stage, the encoder is trained to maximize the mutual information between a set of phonetic targets and corresponding codes. After each stage, the codes are compressed into segments. This step expands acoustic- phonetic context and reduces subsequent computation. We evaluated these MMI encoders by comparing them against a standard minimum distortion (MD) vector quantizer (encoder). Both encoders produced code streams, which were used to train speaker-independent discrete hidden Markov models in a simplified version of the Sphinx system [3]. We used data from the DARPA Resource Management (RM) task. The two-stage cascade of MMI encoders significantly outperforms the standard MD encoder in both speed and accuracy. INTRODUCTION Most hidden Markov model systems use minimum distortion (MD) vector quantizers (encoders) to convert continuously valued speech parameters into streams of integer codes. However, MD encoders do not optimize a criterion that is directly related to recognition accuracy. Moreover, they use a single distortion measure that may not be appropriate for all speech classes. In this paper, we propose the use of maximum mutual information (MMI) encoders that are trained to extract phonetic information and thereby minimize phonetic recognition errors. We further compress the frames into larger segments and repeat the encoding. Our MMI encoders are binary decision trees built to maximize the average mutual information between the phonetic targets and the codes assigned to them. The task of training such encoders has been extensively addressed in the theory of binary decision trees [5, 8, 2]. For example, Breiman et al. systematically consider binary decision trees applied to various classification tasks. The decision (interior) nodes of the tree are allowed to use linear combinations of feature vectors, as well as unordered categorical features. Training criteria ("impurity" criteria) for the binary decision trees include the average leaf-node-conditional class entropy. Training is performed in a top-down node-at-a-time fashion, adding new leaf nodes and maximizing reduction in the average leaf node impurity attained by such additions. It is demonstrated on many practical classification problems that the above procedure results in a suboptimal, but sufficiently accurate tree. Labelled data necessary for the supervised training is obtained by aligning speech frames with phonetic transcriptions using dynamic programming. We train a two- stage cascade of binary-tree encoders. In the first stage, frames are encoded to extract maximum information about their target label classes. Feature vectors used in the tree encoder are frame-based. Contiguous runs of frames with the same code are compressed into segments. In the second stage, the resulting segments are encoded to extract maximum information about their target label classes (we assign a single target label class per …

Content truncated to 500 words