DEVELOPMENT OF NORMALIZATION OF SPEECH FOR AUTOMATIC RECOGNITION

Authors
  1. Lennig, M.
Corporate Authors
Bell-Northern Research Ltd, Ottawa ONT (CAN);Defence and Civil Inst of Environmental Medicine, Downsview ONT (CAN)
Abstract
Vowel recognition experiments are used to compare various preprocessing and normalization techniques. Data consist of a list of ninety French words read by ten Parisian speakers. Vowels of the final stressed syllable of each word are hand segmented. Both speaker-dependent and speaker-independent recognition experiments are carried out on fifteen oral and nasalized vowel phonemes. In addition to recognition accuracy, normalization efficacy is measured using the ratio of between-phoneme to within-phoneme variance. Results show that the mel-based cepstrum, without the overall energy coefficient, performs better than individually or globally determined principal components. The most effective normalization technique is found to be vowel space centering. When eight cepstral coefficients are used, centering raises the speaker-independent recognition score from 68 percent to 80 percent. TRUNCATED
Report Number
BNR-TR-6784 — Final Report
Date of publication
15 Jan 1984
Number of Pages
126
DSTKIM No
84-02696
CANDIS No
91904
Format(s):
Microfiche filmed at DSIS;Originator's fiche received by DSIS;Hardcopy

Permanent link

Document 1 of 1

Date modified: