14th International Congress of Phonetic Sciences (ICPhS-14)

San Francisco, CA, USA
August 1-7, 1999

Converting Speech Signals to Phonological Features

Henning Reetz

Department of Linguistics, University of Konstanz, Germany

The FUL speech recognition system [1] uses underspecified phonological features for the lexical access of words. These features are extracted from the acoustic signal by standard LPCanalysis and by the computation of rough spectral shape parameters. The obtained streams of formants and spectral shape parameters are corrected on a local and global basis by an algorithm that mimics the visual perception of the formant/shape traces. The corrected traces are converted by simple logical decisions into sets of features. Short disturbances are removed from these feature sets and the individual features are time-aligned to give sets of synchronously changing features. These aligned feature sets are used to access possible word candidates from a lexicon of 50 000 base word forms.

Reference

  1. Lahiri, A. 1999. Speech recognition with phonological features. The XIVth International Congress of Phonetic Sciences, San Francisco

Full Paper

Bibliographic reference.  Reetz, Henning (1999): "Converting speech signals to phonological features", In ICPhS-14, 1733-1736.