On The Robustness of Phonetic Information In Short-Time Speech Spectra

Auteurs-es

  • Meg Withgott
  • Marcia A. Bush

Résumé

Speech recognition techniques which take fixed-time slices as input to a matcher face the task of mapping from arbitrary pieces of the physical signal to abstract linguistic units. This paper examines the reliability with which individual vector-quantized LPC spectra can be mapped to various sets of acoustic-phonetic classes. The database for the experiments consisted of approximately 130,000 spectra from a pre-labeled corpus of 616 5-digit strings, and classification was performed on the basis of a maximum likehood decision rule. Classification accuracy, when the same database was used for training and testing, ranged from 94.0% for a simple voiced-voiceless distinction to 42 .7% for a set of 45 acoustic-phonetic classes used in earlier connected digit recognition experiments

Fichiers supplémentaires

Publié-e

2022-12-03

Comment citer

1.
Withgott M, Bush MA. On The Robustness of Phonetic Information In Short-Time Speech Spectra. Canadian Acoustics [Internet]. 3 déc. 2022 [cité 7 août 2024];14(3 bis):101-2. Disponible à: https://jcaa.caa-aca.ca/index.php/jcaa/article/view/3540

Numéro

Rubrique

Actes du congrès de la Semaine canadienne d'acoustique