Dragon Systems Resource Management Benchmark Results February 1991

DRAGON SYSTEMS RESOURCE MANAGEMENT BENCHMARK RESULTS FEBRUARY 19911 James Baker, Janet Baker, Pard Bamberg, Larry Gillick, Lori LameI, Robert Roth, Francesco Scattone, Dean Sturtevant, Ousmane Ba, Richard Benedict Dragon Systems, Inc. 320 Nevada Street Newton, Massachusetts 02160 DRAGON@A.ISI.EDU TEL: (617) 965-5200 FAX: (617) 527-0372 ABSTRACT In this paper we present preliminary results obtained at Dragon Systems on the Resource Maaaagernent benchmark task. The basic conceptual units of our system are Phonemes-m-Context (PICs), which are represented as Hidden Mmkov Models, each of which is eapressed as a sequence of Phonetic Elements (PELs). The PELs corresponding to a given phoneme constitute a kind of alphabet for the representation of PICs. For the speaker-dependent tests, two basic methods of training the acoustic models were investigated. 'nac first method of training the Resouro~ Managemera models is to ~e-estimate the models for each test speaker from that speaker's training data, keeping the PEL spellings of the PICs fixed. The second approach is to use the re-estimated models from the first melhod to derive a segmentation of the training data, then to respall the PICs in a hrgely speaker-depmdmt manner in order to improve the representation of speaker differences. A full explanation of these methods is given, as are results using each method. In addition to repotting on two different training slrategies, we disoass N- Best results. The N-Best algorithm is a modification of the algorithm proposed by Soong and Huang at the Jtme 1990 workshop. This algorithm runs as a post-processing step and uses an A*-search (an algorithm also known as a 'stack decoder'). 1. INTRODUCTION In this paper we report on some preliminary work done at Dragon Systems' on the Resource Management benchmark task. First, a brief overview of Dragon Systems speaker- dependent, continuous speech recognition system is given. Next, the modifications necessary to evaluate this system on the RM task are described. Our goal has been to make changes to the standard continuous speech recognition system in ways that are in line with Dragon's long term aims. The primary modifications so far have been in the areas of signal processing and speaker-dependent training. The speaker-dependent training is described in detail in Section 4. Recognition results are given for the RM1 speaker- dependent development test data and for the Feb91 evaluation test material. In presenting these results, we make a start at evaluating the transfer characteristics of our system when responding to changes in the speaker, the hardware, and the signal processing algorithm. Our experimentation was performed using the speaker-dependent development test ~ta~ and these dam arc used to compare system configurations in this paper. Since we befieve that we are still on a steep learning curve, the February 1991 evaluation lest material was tun through the system only one time, and thus comparative results using the evaluation dam are not yet available. 2. OVERVIEW OF THE DRAGON CSR SYSTEM Dragon Systems' continuous speech recognition system was presented at the June 1990 DARPA meeting [1,2,3]. The system is speaker-dependent and was demonstrated to be …

Content truncated to 500 words