Recognition of the spoken French alphabet using a two-pass dynamic time warp algorithm

Authors

  • J.-P. Cordeau INRS-Telecomm., Verdun, Que., Canada
  • P. Mermelstein INRS-Telecomm., Verdun, Que., Canada

Keywords:

frequency-domain analysis, spectral analysis, speech recognition, time-domain analysis, isolated word recognition, word classes differentiation, confusable sets members discrimination, within-class tokens separation, temporal regions, spoken French alphabet, two-pass dynamic time warp algorithm, acoustically similar words, within-class words, special recognition approach, highly confusable sets, cepstral regions, greatest between-word variance, modified frame-specific weighting scheme, error rate

Abstract

The spoken French alphabet is composed of acoustically similar words that can be organized into six distinct classes. For recognizers working on such a data base, the words not recognized properly are always within-class words and a special recognition approach must be introduced to overcome these effects. A two-pass algorithm is used to discriminate between members of the highly confusable sets. The first pass is used to differentiate similar word classes, while the second pass uses a discriminant analysis to separate within-class tokens. The second stage provides better discrimination, separating the words within each class through improved focus on the temporal and cepstral regions of greatest between-word variance. The improvement in discriminability is provided by a modified frame-specific weighting scheme. The error rate is reduced by 50% using this approach as compared to the direct one-step DTW algorithms

Additional Files

Published

1989-07-01

How to Cite

1.
Cordeau J-P, Mermelstein P. Recognition of the spoken French alphabet using a two-pass dynamic time warp algorithm. Canadian Acoustics [Internet]. 1989 Jul. 1 [cited 2024 May 23];17(3):3-13. Available from: https://jcaa.caa-aca.ca/index.php/jcaa/article/view/608

Issue

Section

Technical Articles