5th International Conference on Spoken Language Processing

Sydney, Australia
November 30 - December 4, 1998

Favourable and Unfavourable Short Duration Segments of Speech in Noise

Daniel Woo

The University of New South Wales, Australia

Human perceptual experiments are described that present listeners with segmented stop consonant speech stimuli in noise. The selection of short duration speech segments is based on a local measure of the signal-to-noise ratio calculated over 1ms windows. The aim is to create stimuli with known fluctuations occurring between a speech and noise sample to assess whether the presence of short duration "gaps" in the noise produce favourable and unfavourable signal regions that influence identification. Perceptual results are reported that suggest human listeners make better use of signals that comprise only of positive, local signal-to-noise ratio segments. Such regions are assumed to be more favourable for stimuli identification. Presentation of stimuli containing only negative signal-to-noise ratio regions does not appear to contribute as much. A model that is based on the accumulation of short duration spectral segments is presented that produces a similar set of identification functions for the same test stimuli.

Full Paper

Bibliographic reference.  Woo, Daniel (1998): "Favourable and unfavourable short duration segments of speech in noise", In ICSLP-1998, paper 0700.