Re: [vmeet] Remote Attendance: Simultaneous Translation Software

[John C Klensin <[e-mail]>]

--On Saturday, April 09, 2016 14:03 +0000 "Alvaro Retana
(aretana)" <[e-mail]> wrote:

> If it is regarding interpretation, I think the first step
> would be - good speech to text (english), with a possible
> input to etherpad like tool, that could be corrected by minute
> taker.
> 
> I really like Dhruv's suggestion.  The result would help even
> the native English speakers and wouldn't be limited to remote
> participants.

Folks, fwiw, I've run IETF sessions through good-quality speech
to text programs.  The results have not been very helpful even
when the systems have been trained with dictionaries built by
scanning most of the RFCs I've written and a selection of
others.  When run with generic voices (rather than training to
my peculiar voice and speech patterns), the systems basically go
nuts and become useless when confronted with heavily-accented
English or "unusual" sentence rhythms.   Even those
unsatisfactory results aren't real-time either -- the systems
are reasonable for recording speech and delivering text with
some considerable, and cumulative, lag.

One could improve on that with more powerful software,
dictionaries, and computers, but, for example, I don't think
IASA or AMS have a supply of free supercomputer cycles.   The
ICANN transcription efforts, at least when I was last paying a
lot of attention, were different because they involved highly
skilled human transcribers and, at least as important, the
fraction of the ICANN discussions that involve highly technical
material tends to be quite small.

I'd be happy to have written transcripts too and I'm a native
English speaker.  However, please, let's try to keep this real
and, if we have expensive ideas, to be sure we understand where
the resources might come from and what would have to be pushed
to lower priority.

    john