Spectro-Temporal Modulation Transfer Functions and Speech Intelligibility.

摘要

The most obvious feature of a speech spectrogram is the energy modulations, both in time in any given frequency channel, and along the spectral axis at any instant, due to formant peaks and their transitions, spectral edges, and rapid amplitude modulations at onsets /offsets. These modulations occur at relatively slow temporal rates (few Hz) reflecting the speed of the articulatory gestures, and hence the phonetic and syllabic rates of speech. Speech intelligibility is critically dependent on the clarity of these spectro-temporal modulations. Thus speech reconstructed from smoothed spectrograms along either dimension suffers from progressive loss of intelligibility.