[Corpora-List] Corpus methods for prosodic analysis

Dave Graff graff at ldc.upenn.edu
Wed Apr 16 16:43:29 CEST 2014


Matías,

Assuming that you have digital speech recordings and the transcripts derived from them, but the transcripts do not have (adequate) time stamps, a sensible first step would be to apply "forced alignment" of the transcripts to the speech, so that you get word- and phoneme-level time stamps applied to the transcripts. But that requires having a fairly good set of acoustic models for the language in question. (For English, look at http://www.ling.upenn.edu/phonetics/p2fa/)

As for doing any "automatic" check for prosodic patterns, that's a difficult matter, since prosody is a complex interaction of timing and intensity as well as F0. My first inclination would be to reduce the question to a relatively simple set of alternations that can be described clearly, and then manually audit the relevant bits of audio corresponding to the relevant bits of transcript, marking each instance by selecting from the fixed set of (say, two to five) alternative labels.

Having adequate time stamps in the transcript (i.e. at least at the level of phrases) will speed that up. The main point is that in order to use numerical/statistical analysis to label prosodic events, manual labeling has to be done by humans first in any case, to train the analysis engine.

Best regards,

David Graff

On Apr 16, 2014, at 2:35 AM, Matías Guzmán Naranjo <mortem.dei at gmail.com> wrote:


> Dear all,
>
> I found a strong correlation between what seems are sentence boundaries (in this case marked as punctuation marks), and a particular morphological alternation. Now I want to actually check for prosodic patterns in an oral corpus and see whether the correlation is actually there or if it's just because of idiosyncrasies of how the transcriber annotated the corpus.
>
> My question is this, are there any techniques for doing this automatically? Something like say, manually splitting the audio files in the relevant text units, but automatically doing an F0 analysis, or something similar. Also, is there any previous research that does this?
>
> Thanks for your help,
>
> Matías
> _______________________________________________
> UNSUBSCRIBE from this page: http://mailman.uib.no/options/corpora
> Corpora mailing list
> Corpora at uib.no
> http://mailman.uib.no/listinfo/corpora



More information about the Corpora mailing list