Coding of the speech spectrum in three time-varying sinusoids

Recent perceptual experiments with normal adult listeners show that phonetic information can readily be conveyed by sinewave replicas of speech signals. These tonal patterns are made of three sinusoids set equal in frequency and amplitude to the respective peaks of the first three formants of natural-speech utterances. Unlike natural and most synthetic speech, the spectrum of sinusoidal patterns contains neither harmonics nor broadband formants, and is identified as grossly unnatural in voice timbre. Despite this drastic recoding of the short-time speech spectrum, listeners perceive the phonetic content if the temporal properties of spectrum variation are preserved. These observations suggest that phonetic perception may depend on properties of coherent spectrum variation, a second-order property of the acoustic signal, rather than any particular set of acoustic elements present in speech signals.