I think it would be fair to say that there are mixed opinions about
UniVoice in the voice recognition community. My information is that it
works best with US English pronunciation and that while it is okay-ish
for dictation (ie single user, using a decent microphone) for a mixture
of voices, and particularly for multiple voices in conversation it is of
questionable value. If anybody hears any different, I would be really
interested to know since I seem to spend an awful lot of time doing
transcription at the moment....
David Cairns
Dr David Cairns, Assistant Director, Development and Enhancement Group,
QAA
Southgate House, Southgate Street, Gloucester, GL1 1UB
Tel: 01452 55 7106 (direct line) Fax: 01452 55 7011
Mobile: 07774 816875
PA: Rachel Curtin ([log in to unmask])
Tel: 01452 55 7138 (direct line)
This message has been dictated using Dragon NaturallySpeaking speech to
text software. Some homophones (there, their etc) may be
misrecognised...
-----Original Message-----
From: qual-software [mailto:[log in to unmask]] On Behalf Of
Duncan Branley
Sent: 20 March 2007 17:08
To: [log in to unmask]
Subject: Voice Recognition of Multiple Voices: UniVoice
While looking at some information about Dragon Naturally Speaking today
I came across an add-on which claims to be able to transcribe 4 voices
as part of the same recording without needing separate profiles for each
voice. It's part of KnowBrainer and is called UniVoice. I wonder has
anyone used it? It was released last autumn.
<https://www.knowbrainer.com/storefront/index.cfm?fuseaction=category.di
splay&category_id=19&CFID=686607&CFTOKEN=21813192&jsessionid=a0301948511
174408439795>
This could be incredibly useful for transcribing interviews and focus
groups if the number could be increased.
There's a bit of text on the forum (to which you have to subscribe to
read):
"Having 4 people read a hundred words each into NaturallySpeaking hardly
constitutes a serious test but we were not looking for a weekend
project.
We simply wanted to perform some preliminary tests of UniVoice to see
how it stacks up and we have concluded the following: This variation of
UniVoice is a highly accurate non-gender specific pre-trained user
profile.
Mark, uses NaturallySpeaking regularly and averages about 98% accuracy
when reading text. Mark's first test was only 95% accurate. When
reading the same text in UniVoice he repeatedly scored 100% accuracy.
Mark produced a 2% gain in accuracy which translates into 50% fewer
errors to correct. In Mark's case there were no errors to correct but
keep in mind that he was reading text. In real-life conditions the
accuracy of both UniVoice and a trained user file will be lower. A 50%
reduction in error correcting is a huge leap in speech recognition
technology. "
<http://www.knowbrainer.com/PubForum/index.cfm?page=viewForumTopic&topic
Id=41>
Duncan
===================================================
Duncan Branley [log in to unmask]
Research Applications Officer, I T Services
Goldsmiths, University of London
New Cross, LONDON SE14 6NW
Tel: +44 (0)20 7919 7708 Fax: +44 (0)20 7919 7556
===================================================
**********
If you receive this message in error, please inform the sender and delete the message.
All contents have been scanned by QAA, although this is not a guarantee that they are virus free.
The Quality Assurance Agency for Higher Education is a company limited by guarantee.
Registered in England and Wales (No.3344784)
Registered office Southgate House, Southgate Street, Gloucester GL1 1UB.
Registered charity no. 1062746
A charity registered in Scotland no. SC037786
|