Perception of atomic speech

Abstract

Natural speech signals are continuous and have great redundancy in time, which is learned by human beings and guarantees speech comprehension even in challenging circumstances. This work develops a sound synthesis method to study some time-related characteristics of the human auditory system. The temporal information within the individual frequency band of a speech is down-sampled using Gaussian-shaped pulses and then recombined into a new sound which may have some remaining intelligibility or feasible phonetic features. The “atomic” sound is coined to the sound with an extremely spectral-temporally sparse spectrogram generated using the method. A battery of speech perception tests was administered in normal-hearing listeners. Results show that (1) atomic sounds from clear speech can be understood as speech, although the listeners often reported a feeling of water sound textures; (2) the temporal and spectral resolution could be traded off in the atomic speech comprehension; and (3) only one maximum envelope value preserved among a 32-channel filter-bank at a rate of 400 Hz was surprisingly adequate for speech understanding, which indicates that the brain can organize the only spectral peak within each short duration no longer than 2.5-ms into sentence understanding without either explicit or implicit encoding of the first three formants.