Automatic measurement of speech recognition performance: a comparison of six speaker-dependent recognition devices

Howard C. Nusbaum, David B. Pisoni

Research output: Contribution to journalArticle

6 Scopus citations

Abstract

Although performance data are often freely cited by vendors of speech recognition devices, the conditions under which these data were collected are seldom specified in sufficient detail to permit comparisons among different systems. To directly compare the performance of six commercially available speech recognition devices, we developed a computer-controlled testing system and a set of standard tests. We carried out these tests to assess the performance of speech recognition devices sold by Texas Instruments, Votan, Dragon, IBM, Interstate, and NEC. The results demonstrate several reliable performance differences among these systems. In general, however, performance differences among these devices are quite small and are reduced by appropriate training. Our results also indicate that the effects of training on performance of a speech recognition device are much greater for difficult vocabularies than they are for discriminable vocabularies. Finally, an examination of the results for recognition of the speech produced by one talker in the testing database suggests that user-specific difficulties in recognition performance may, in some cases, result from interactions among the application vocabulary, the user's speech, and the algorithm of a recognition device.

Original languageEnglish (US)
Pages (from-to)87-108
Number of pages22
JournalComputer Speech and Language
Volume2
Issue number2
DOIs
StatePublished - Jun 1987

ASJC Scopus subject areas

  • Software
  • Theoretical Computer Science
  • Human-Computer Interaction

Fingerprint Dive into the research topics of 'Automatic measurement of speech recognition performance: a comparison of six speaker-dependent recognition devices'. Together they form a unique fingerprint.

  • Cite this