Re: Read Dialogue Corpus

Bill Fisher (william.fisher@nist.gov)
Tue, 6 May 1997 09:26:42 -0400

On May 6, 1:23pm, Maria Wolters wrote:
> Subject: Read Dialogue Corpus
>
> We are looking for a corpus of phonetically transcribed
> read dialogues, language: American English. Our aim is to identify
> a typical prosody for certain dialogue acts, which will
> then be incorporated into a speech synthesis system.
>
> Spontaneous speech is less of an option because of its
> variability, which could make things difficult for a machine learning
> algorithm.
>

I don't think studying read dialogues will tell you much
about typical prosody for certain dialogue acts. The trouble
is that most people, who are not actors, will put a very
dull and inappropriate prosody on read speech of any kind.
Typically, it's just a list of words to them; it's not a
speech act of any kind.

The best source for studying real prosody in dialogue that
I know of is the "Switchboard" and/or "CallHome" corpora that
are being made available through the Linguistic Data Consortium
at Penn. They are real telephone calls, some between friends
and relatives and some between strangers, and they come with
orthographic transcriptions. In Switchboard at least, time marks
are available for each word, and it's not too hard a job to
do a forced alignment with dictionary pronunciations of the
words to get a phonetic/phonemic transcription; it's been
done at least once that I know of.

Another option you might try is to purchase transcriptions
of television or radio shows that you can record off the air.
For instance, Wall St. Week with Louis Ruykeyser offers
transcripts for each 1/2 hour show for $6. These are real
people who are also experienced show performers, discussing
financial issues among themselves. These are orthographic
transcriptions, not phonetic, but you always have the
option of forced alignment using a speech recognizer to
get a phonetic transcription.

I know real speech is more variable, but that's where the
truth is. Don't look for your keys under the light just
because that's where the seeing is better.

- Bill Fisher