European Conference on Speech Technology Edinburgh, Scotland, UK September 1987

The Use of Rough Spectral Features for Large Vocabulary Recognition

Gilles Adda, Maxine Eskénazi, P. E. Stem

LIMSI - CNRS, Orsay, France

This paper presents a study of the use of rough spectral features which may be detected in the speech signal with a high degree of certainty, for large vocabulary 015,000 words) isolated word recognition. The use of these features must permit correct preclassification, a finer recognition process being used afterward on the vocabulary subset. For correct cohort (class) access, it is essential to propose several feature strings for any given word in order to take into account phonological variability, and it is necessary to compact several occurrences of the same feature into one to avoid segmentation-based errors. We use context-dependant rewrite rules to transform a syllable in phonetic form into one or several strings of features; an engine applies these rules to whole words, then determining the feature string labelled cohorts.
The study was carried out on a 17,000-form vocabulary with confirmation on a 270,000-form one. Maximum, mean, and expected cohort sizes have been calculated. The definition of expected size has been generalised in order to take into account the multiple strings for a given word. The evaluation shows, for the French language, that the use of rough spectral features is interesting, but less vocabulary-reducing, than it seems in other studies, due to the multiple string representation.

Full Paper

Bibliographic reference.  Adda, Gilles / Eskénazi, Maxine / Stem, P. E. (1987): "The use of rough spectral features for large vocabulary recognition", In ECST-1987, 1171-1174.