Time structure and detection of the multivoiced segments in mixed speech

Jean-Sylvain Liénard (1) , Claude Barras (1) , François Signol (1)

1 LIMSI - Laboratoire d'Informatique pour la Mécanique et les Sciences de l'Ingénieur

When two speech signals are mixed in a single channel the voiced parts of any of them remain mostly unaltered during the voicing interruptions of the other, i.e. pauses and voiceless consonants. The mixture is made of 3 types of multivoiced segments noted 0V (unvoiced), 1V (one voicing) and 2V (two voicings). A statistical study of read-aloud texts reveals that total time spent in the 1V state is twice as long as the time spent in any of the other states. The HSC multipitch algorithm, based on a specific mechanism that eliminates the f0 halving and doubling errors, is used to locate the 3 segments types in the signal. This feature is illustrated by the task of spotting a short utterance repeatedly mixed with a long text.

Mots clés: voicing; pitch; multipitch; speech separation

Domaines: Informatique [cs]; Informatique et langage [cs.CL]

Jean-Sylvain Liénard, Claude Barras, François Signol. Time structure and detection of the multivoiced segments in mixed speech. International Congress of Phonetic Sciences, Jan 2011, Hong Kong, China. ⟨hal-01836479⟩