wintermute-devel team mailing list archive

[Jacky Alcine <jacky.alcine@xxxxxxxxxx>]

I've noticed something regarding detecting the WER (Word Error Rate)
or the error rate of the recognition from PocketSphinx. On their Wiki[1], the
formula for the Word Error Rate would be described as follows:

"Let['s say] we have [some] original text and [a] recognition text of
length of *N* words. From [that set], *I* words were inserted; *D*
words were deleted and *S* words were substituted. [The] word error
rate [would be]

WER = (I + D + S) / N

WER is usually measured in percent."

Something to note, is there a means of pulling any of the values of
these variables from the API?

There's also a formula for accuracy and speed but WER seems like the
best way to check this stuff.

[1] http://cmusphinx.sourceforge.net/wiki/tutorialconcepts#what_is_optimized

> On Sun, Dec 11, 2011 at 6:57 PM, Islam F. Bahnasy <ifarouk@xxxxxxxxx> wrote:
>>
>> Hi all,
>>
>> I was able to recognize recorded voice with PocketSphinx as an experiment.
>> We need to make that continuous recognition like dictation so that's the coming part and of course integrate all that with Wintermute.
>>
>> --
>> Sincerily,
>> Islam Bahnasy
>>
>>
>>
>> _______________________________________________
>> Mailing list: https://launchpad.net/~wintermute-devel
>> Post to     : wintermute-devel@xxxxxxxxxxxxxxxxxxx
>> Unsubscribe : https://launchpad.net/~wintermute-devel
>> More help   : https://help.launchpad.net/ListHelp
>>

One Vision. One Purpose.
We are the Synthetic Intellect Institute
contact@xxxxxxxxxx :: http://www.thesii.org/ :: #sii on FreeNode

Jacky Alcine, Head Developer of Wintermute
http://jackyalcine.co.cc :: jacky.alcine@xxxxxxxxxx ::
https://launchpad.net/~jackyalcine