BYBLOS SPEECH RECOGNITION BENCHMARK RESULTS

F. Kubala; S. Austin; C. Barry; J. Makhoul; P. Placeway; R. Schwartz

Description

a speakers. Surprisingly, • (ATIS) 2rid a 1st ATIS features: speakers 45 grammar. • first S. dialects. Austin, DARPA evaluation: DARPA character- conditions: DESCRIPTIOPN a Makhoul, Travel Finally, MIT/Lincoln. evaluation. (SI) domain. a first speakers. BBN speakers. Systems Gaussian individual training. ATIS • ATIS 2. grammar. SI-109 a data. training Travel a conceentrated A (ATIS) 1. 25%. ATIS a statistical statistical acoustic available a a a N--best domain, first carefully ATIS system. adaptation actual speakers simulations 109 all, 12 a 91 BYBLOS SPEECH RECOGNITION BENCHMARK RESULTS F. Kubala, C. Barry, J. P. Placeway, R. Schwartz ABSTRACT lem in several respects. First of and most importantly, the speech was collected during of use of the ATIS This paper presents speech recognition test results from the BBN The speakers were completely uncoached, and therefore, the range of BYBLOS system on the Feb benchmarks in both the speech phenomena goes far beyond that of the controlled Resource Management (RM) and the Air Information System read-speech conditions that exist in the RM corpus. We will describe domains. In the RM test, we report on speaker-independent our recent efforts to deal with these new problems. recognition performance for the standard training condition using speakers and for our recently proposed SI model made from only Since is the ultimate goal of the ATIS we ...

Informations

• Publié par: Chion
• Nombre de lectures: 12
• Langue: English

Extrait