ISCA Archive ICSLP 1994
ISCA Archive ICSLP 1994

Strategies for voice separation based on harmonicity

Alain de Cheveigné

A harmonic sound has a spectrum made of discrete components forming a harmonic series, and its wave form is periodic. Voiced speech is approximately harmonic, and this cue appears to be exploited by the auditory system for the perception of speech in a noisy background. This paper addresses the following question: does the auditory system exploit the harmonic structure of a target to segregate it from the background, or else that of a harmonic background to eliminate it.

The issue was considered from two angles. First of all, both strategies were implemented to reduce voiced interference in a speech recognition experiment. The aim was to determine whether one "works better" than the other in this task; the results showed that harmonic cancellation was more effective. In a second experiment human listeners were presented with pairs of vowels, each of which was either harmonic or inharmonic, and requested to identify them both. The responses were scored according to whether the target vowel was harmonic, and whether the other (interfering) vowel was harmonic. It appears that the auditory system exploits harmonicity of the ground (cancellation strategy), a result that is coherent with the outcome of the first experiment.

Cite as: Cheveigné, A.d. (1994) Strategies for voice separation based on harmonicity. Proc. 3rd International Conference on Spoken Language Processing (ICSLP 1994), 1071-1074

  author={Alain de Cheveigné},
  title={{Strategies for voice separation based on harmonicity}},
  booktitle={Proc. 3rd International Conference on Spoken Language Processing (ICSLP 1994)},