Comparative Experiments on Large Vocabulary Speech Recognition

COMPARATIVE EXPERIMENTS ON LARGE VOCABULARY SPEECH RECOGNITION Richard Schwartz, Tasos Anastasakos, Francis Kubala, John Makhoul, Long Nguyen, George Zavaliagkos BBN Systems & Technologies 70 Fawcett Street. Cambridge, MA 02138 ABSTRACT This paper describes several key experiments in large vocabu- lay speech recognition. We demonstrate that, counter to our intuitions, given a fixed amount of training speech, the num- ber of training speakers has little effect on the accuracy. We show how much speech is needed for speaker-independent (SI) recognition in order to achieve the same performance as speaker- dependent (SD) recognition. We demonstrate that, though the N-Best Paradigm works quite well up to vocabularies of 5,000 words, it begins to break down with 20,000 words and long sen- tences. We compare the performance of two feature preprocess- ing algorithms for microphone independence and we describe a new microphone adaptation algorithm based on selection among several codebook transformations. 1. INTRODUCTION During the past year, the DARPA program has graduated from medium vocabulary recognition problems like Resource Manage- ment and ATIS into the large vocabulary dictation of Wall Street Joumal (WSJ) texts. With this move comes some changes in com- putational requirements and the possibility that the algorithms that worked best on smaller vocabularies would not be the same ones that work best on larger vocabularies. We found that, while the required computation certainly increased, the programs that we had developed on the smaller problems still worked efficiently enough on the larger problems. However, while the BYBLOS system achieved the lowest word error rate obtained by any site for recognition of ATIS speech, the error rates for the WSJ tests were the second lowest of the six sites that tested their systems on this corpus. The reader will find more details on the evaluation results in [1]. In the sections that follow, we will describe the BBN BYBLOS system briefly. Then we enumerate several modifications to the BBN BYBLOS system. Following this we will describe four different experiments that we performed and the results obtained. 2. BYBLOS All of the experiments that will be described were performed using the BBN BYBLOS speech recognition system. This sys- tem introduced an effective strategy for using context-dependent phonetic hidden Markov models (HMM) and demonstrated their feasibility for large vocabulary, continuous speech applications [2]. Over the years, the core algorithms have been refined with improved algorithms for estimating robust speech models and us- ing them effectively to search for the most likely sentence. The system can be trained using the pooled speech of many speakers or by training separate models for each speaker and then averaging the resulting models. The system can be constrained by any finite-state language model, which includes probabilisfic n-gram models as a special case. Nonfinite-state models can also be used in a post process through the N-best Paradigm. The BYBLOS speech recognition system uses a multi-pass search strategy designed to use progressively more detailed mod- els on a correspondingly reduced search space. It produces an ordered list of the N top-scoring hypotheses …

Content truncated to 500 words