VoxForge
VoxForge
Given that the voxforge acoustic models training process uses the CMU dictionary (of American English), how compatible is the result British English and other accents, and what sort of adjustments do I need to make to my inputs (e.g. transcriptions) in order to use the models in such cases?
BTW, I see that the models were trained with diverse speakers and accents, not just American English speakers. Why is this not a source of error bearing in mind that the training process uses the CMU dictionary.
Please note that I'm not criticizing the approach (I don't know enough to do so), I'm just trying to understand.
Thanks.
--- (Edited on 7/25/2013 7:06 pm [GMT-0500] by Olumide) ---
> Given that the voxforge acoustic models training process uses the CMU dictionary (of American English), how compatible is the result British English and other accents
What do you mean by "compatible"?
> and what sort of adjustments do I need to make to my inputs (e.g. transcriptions) in order to use the models in such cases?
> BTW, I see that the models were trained with diverse speakers and accents, not just American English speakers. Why is this not a source of error bearing in mind that the training process uses the CMU dictionary.
This is a source for speech recognition errors on native speech. Error is compensated a little because gaussian mixture model can learn sound diversity including accented speech patterns but it can not learn it properly so some degradation on native speech is present.
--- (Edited on 7/26/2013 10:54 [GMT+0400] by nsh) ---
I mean how compatible is an acoutic model that was trained using an American English dictionary on British English input speech, and what sort of adjustments should be made to the corresponding input (British English) transcription when performing forced alignment?
--- (Edited on 7/26/2013 5:28 am [GMT-0500] by Olumide) ---
> I mean how compatible is an acoutic model that was trained using an American English dictionary on British English input speech
I still not sure what do you mean by "how compatible". Do you expect some number as an answer or what?
> and what sort of adjustments should be made to the corresponding input (British English) transcription when performing forced alignment?
In such a problem statement the only reasonable adjustment is to use British model instead of US one.
--- (Edited on 7/27/2013 01:20 [GMT+0400] by nsh) ---
I found some information on the CMU Sphinx site that cleared things up somewhat.
On adapting anacoustic model, taken from http://cmusphinx.sourceforge.net/wiki/tutorialadapt
Cross-language adaptation also make sense, for example you can adapt English model to sounds of other language by creating a phoneset map and creating other language dictionary with English phoneset.
--- (Edited on 7/27/2013 8:22 pm [GMT-0500] by Olumide) ---
Hi,
Do you now how to create a phoneset map to adapt acoustic model with other language dictionary such as English phoneset? I have found beep dictionary a british dictionary but i am not able to use it because acoustic model was trained on english us phoneset.
Thanks in advance,
gf
--- (Edited on 11/13/2013 2:04 am [GMT-0600] by Visitor) ---
To map Beep phoneset to US English phoneset you can just open beep phoneset and write down the corresponding US phone for every phone. Mapping is straightforward if you know which sounds phone represents, most of the phones are direct equivalents:
B -> B
T -> T
IH -> IH
the only issue is to map phone AX from Beep:
AX -> AH
Once you made this map you can convert beep lexicon from UK phonset to US phoneset and adapt the model using this converted lexicon.
--- (Edited on 11/18/2013 03:03 [GMT+0300] by nsh) ---