BYBLOS Speech Recognition Benchmark Results

BYBLOS SPEECH RECOGNITION BENCHMARK RESULTS F. Kubala, S. Austin, C. Barry, J. Makhoul, P. Placeway, R. Schwartz BBN Systems and Technologies Cambridge MA 02138 ABSTRACT This paper presents speech recognition test results from the BBN BYBLOS system on the Feb 91 DARPA benchmarks in both the Resource Management (RM) and the Air Travel Information System (ATIS) domains. In the RM test, we report on speaker-independent (SI) recognition performance for the standard training condition using 109 speakers and for our recently proposed SI model made from only 12 training speakers. Surprisingly, the 12-speaker model performs as well as the one made from 109 speakers. Also within the RM do- main, we demonstrate that state-of-the-art SI models perform poorly for speakers with strong dialects. But we show that this degradation can be overcome by using speaker adaptation from multiple-reference speakers. For the ATIS benchmarks, we ran a new system conligu- ration which first produced a rank-ordered list of the N--best word- sequence hypotheses. The list of hypotheses was then reordered using more detailed acoustic and language models. In the ATIS bench- marks, we report SI recognition results on two conditions. The first is a baseline condition using only training data available from NIST on CD-ROM and a word-based statistical hi-gram grammar developed at MIT/Lincoln. In the second condition, we added training data from speakers collected at BBN and used a 4-gram class grammar. These changes reduced the word error rate by 25%. INTRODUCTION This paper will present new test results for running the BBN BYBLOS system on the speech recognition benchmarks in both the Resource Management (RM) and the Air Travel Information System (ATIS) domains. During this reporting period we have conceentrated on speaker- independent recognition conditions. However, we will also report a new result demonstrating the need and usefulness of speaker adap- tation in order to be able to recognize the speech of speakers with different dialects than those found in the training data. For the RM corpus, we report on three conditions: 1. The common SI-109 training condition that has been widely reported in the past. 2. The new SI-12 training paradigm that we introduced at the previous DARPA workshop. 3. Adaptation to the dialect of the speaker The ATIS domain presents a new type of speech recognition prob- lem in several respects. First of all, and most importantly, the speech was collected during simulations of actual use of the ATIS system. The speakers were completely uncoached, and therefore, the range of speech phenomena goes far beyond that of the carefully controlled read-speech conditions that exist in the RM corpus. We will describe our recent efforts to deal with these new problems. Since understanding is the ultimate goal of the ATIS domain, we use a rank ordered list of the N-best speech recognition hypotheses as the interface to the natural language component of BBN's spoken language system. Below, we desoribe a new procedure which allows the system to use powerful but eomputationally prohibitive acoustic models and statistical grammars to reorder …

Content truncated to 500 words