LINGUIST List 15.2363

Tue Aug 24 2004

Sum: Speech Corpus for Neural Network Training

Editor for this issue: Megan Zdrojkowsky <meganlinguistlist.org>

Directory

  • Scott Drellishak, Speech Corpus for Neural Network Training

    Message 1: Speech Corpus for Neural Network Training

     Date: Mon, 23 Aug 2004 21:18:19 -0400 (EDT)
    From: Scott Drellishak <sfdu.washington.edu>
    Subject: Speech Corpus for Neural Network Training


    A few weeks ago, I posted a request for information about speech corpora of a particular kind to both the Linguist List and the Corpora-List (Linguist 15.1895). This is the (somewhat belated) summary.

    I described the corpora we are seeking as follows:

    ''We are looking for a corpus that contains samples of many speakers producing many vowels (preferably in a less reduced register) that also contains human-validated pitch and formant (F1, F2, and F3) tracks and, if possible, bandwidth information. A corpus that contains more than just vowels is fine, since we can discard sections of the samples that do not suit our needs.''

    I received five replies:

    1) John Lawler suggested MICASE (Michigan Corpus of Academic Spoken English), which is available here:

    http://www.lsa.umich.edu/eli/micase/index.htm

    2) Lesley Carmichael suggested I post my request to the Corpora-List.

    3) Jane Edwards pointed me at the Switchboard Transcription Project:

    http://www.icsi.berkeley.edu/real/stp/index.html

    4) Susana Sotillo wrote, ''At a recent conference (CALICO) I saw a demonstration of the Speechcalator (Allen Blackwell and associates). Why don't you write him at Carnegie- Mellon.''

    5) Linda Bawcom offered an hour and a half of taped conversation that she used in her MA research.

    Many thanks to everyone who replied.

    Scott Drellishak University of Washington Seattle, WA