Speech perception from a crudely quantized spectrogram: A figure‐ground analogy

Abstract

Recent research into auditory chimeras has shown that comprehensible speech can be generated by superimposing the frequency‐dependent modulation envelope of a target speech signal onto the fine structure of music, speech, or almost any other broadband waveform. In this experiment, we show that, in certain cases, intelligibility can be achieved with an even simpler stimulus that preserves only the rough outline of the energy distribution of the target speech waveform. This stimulus was generated by 1) dividing the target speech spectrogram into discrete tiles that were 1/3rd octave band wide and 7.8 ms long; 2) identifying the minimum number of tiles necessary to capture 90% of the energy in the target waveform; and 3) filtering a uniform, broadband noise to activate only those tiles that met the threshold energy criterion in the original stimulus. The resulting stimulus captures the overall outline of the distribution of energy in the target speech, but contains no information about the relative distribution of energy within the spectrogram. In this paper, we discuss the use of this stimulus in a series of auditory figure‐ground experiments where the target speech signal differs from the background by only a single parameter (relative level, ITD, ILD, etc.).