Implementation Aspects of Large Vocabulary Recognition Based on Intraword and Interword Phonetic Units

IMPLEMENTATION ASPECTS OF LARGE VOCABULARY RECOGNITION BASED ON INTRAWORD AND INTERWORD PHONETIC UNITS R. Pieraccini, C. H. Lee, E. Giachint , L. R. Rabiner Speech Research Department AT&T Bell Laboratories Murray Hill, NJ 07974 ABSTRACT Most large vocabulary speech recognition systems essentially consist of a training algorithm and a recognition structure which is essentially a search for the best path through a rather large decoding network. Although the performance of the recognizer is crucially tied to the details of the training procedure, it is absolutely essential that the recognition structure be efficient in terms of computation and memory, and accurate in terms of actually determining the best path through the lattice, so that a wide range of training (sub- word unit creation) strategies can be efficiently evaluated in a reasonable time period. We have considered an architecture in which we incorporate several well known procedures (beam search, compiled network, etc.) with some new ideas (stacks of active network nodes, likelihood computation on demand, guided search, etc.) to implement a search procedure which maintains the accuracy of the full search but which can decode a single sentence in about one minute of computing time (about 20 times real time) on a vectorized, concurrent processor. The ways in which we have realized this significant computational reduction are described in this paper. I. INTRODUCTION Most large vocabulary speech recognition systems are implemented as network searches for the best path through a large, but finite grid. The best path generally corresponds to the most likely sequence of words as constrained by a finite state network which implements the grammar or syntactic component of the system. When the number of basic speech (sub-word) units is small (i.e., on the order of 50-200 units), the details of implementation of the search strategy don't have a major impact on the overall complexity of the recognition task and experimental tuning and assessment of different training strategies and overall performance of the recognizer is relatively straightforward. However, when t Now with CSELT, Torino, Italy. highly detailed (context dependent) speech units are used, including both intraword and interword context dependent units, the complexity of the overall implementation often increases quadratically with the number of basic units, and correspondingly the details of how the network search is implemented become of major importance in accessing the suitability of various network structures. We have especially found this to be the case when we use context dependent interword speech units where the fan-out at the end of a word and the fan-in at the beginning of the next word are both very high, and the bookkeeping to keep track of all possible phonetic contexts in which a word occurs can dominate the computation (to the extent that it can make even the likelihood computation with mixture density HMM's be negligible compared to the bookkeeping overhead), since at every frame, all possible connections between every valid pair of words in the grammar must be examined. When this is the case, a full search implementation of …

Content truncated to 500 words