Subject: Re: musical tones in speech From: "Alain de Cheveigne'" <Alain.de.Cheveigne(at)IRCAM.FR> Date: Sat, 12 May 2001 13:22:13 +0200Bruno, >How do you respond to Martin's point (with which I agree) that there >is no hope of obtaining any distribution with discrete peaks from >quasi-continuous F0 contours? Such contours consist in large part of >the transitions between the inflection points that are believed to be >significant in intonation. If there is no hope of obtaining a distribution with discrete peaks from quasi-continous F0 contours, you should conclude that the discrete distribution reported by Braun is artefactual. Remember that speech targets are supposed to have been derived (manually) from F0 contours. I verified, for another database, that these contours themselves show no discrete peaks in their distribution. If the distribution reported by Braun is not artefactual, it follows that there IS a way of obtaining a distribution with discrete peaks from quasi-continuous F0 contours, in contradiction with your and Martin's point. I tested one plausible way. I looked at the populations of minima and maxima of contiguous voiced portions, among which I'd expect to find likely candidates for "speech targets". Manual selection of target points might reject some of these candidates and add others, but I'd expect a special distribution of speech targets, if it were real, to be reflected by a special distribution of these minima or maxima. I found none. I can't rule out the possibility that the algorithm used by the hand-markers was more sophisticated, capable of producing musical-scale-like distributions where simpler algorithms cannot. Unfortunately the paper gives no information beyond a reference that is not as easily accessible as ARLO. Perhaps a useful next step would be for Martin to describe how speech targets were obtained. I would like sufficient detail to implement an automatic process by which I can produce targets, not necessarily as reliable as those produced by hand, but that are likely to show a roughly similar distribution. If automatic marking is not possible, then please explain why, and what aspect specific to hand-marking produced the remarkable distribution. Then perhaps I'll feel comfortable that it exists. Alain -------------------------------------------------------------- Alain de Cheveigne' CNRS/IRCAM, 1 place Stravinsky, 75004, Paris. phone: +33 1 44784846, fax: 44781540, email: cheveign(at)ircam.fr http://www.ircam.fr/equipes/pcm/cheveign --------------------------------------------------------------