Hi,
Does anyone know if a serious attempt at building a German acoustic model has been done by anyone. In searching the archives I've seen one or two similar inquiries, but there were no indications as to whether anything was continued.
Also, just how difficult is it to build up a new acoustic model. I have no experience with what would be required to do this, but I am interested to know how long it might take a computer savi German linguist to do.
many thanks
robert
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
2. My understanding of Sphinx is, that one should not try to use it as a replacement for something like ViaVoice, but rather as a very flexible open source tool for building speakerindependent speech controled applications.
For example, I will try to implement it with a whole bunch of language models. One model for each menue and none with more than just some hundred words, which are a mixture of German, English, Latin, and product names. That is to say, I will build one model for each database column and entryfield, since this seems to be the best way to maximize the recognition rate, speakerindependence and speed.
If seems, that with up to 50 words one should train just the words as phonems, while with more words one should try to make a cleverly tailored selection of phonems (in German 'Silben').
A good explantion to start with seems to be man1.html at the SphinxTrain page.
A complete German model only seems to be usefull if one wishes to use Sphinx with rather large vocabularies.
Though the strength of Sphinx seems to be rather the ability to limit the vocabulary to what is really needed, and to switch quickly between language models.
However, I have just studied somewhat the manuals but can not yet play with Sphinx since SphinxTrain does not compile on Suse 8.2's gcc-3.3. Do you have a solution to this?
Regards,
Christoph
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi,
Thanks for the reply. At the minute I do only need a small vocabulary control application, so Sphinx should do me as it is.
However I still feel that a serious open source replacement for ViaVoice is required. The people over at XVoice seem to think Sphinx is that replacement for the time being. If the feeling is that Sphinx will never scale up to full dictation, then maybe they need to consider another replacement for ViaVoice.
I have not yet made any attempt to use SphinxTrain, I'm still in the stages of asking around to see what has been done. Also I have no training in linguistics, so I'm not really sure of what I would have to do anyway. I might yet come back to use SphinxTrain.
How much actual time effort do you think it would take to build a german acoustic model which would be complex enough for the Turtle example?
yours gratefully
robert
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi,
Does anyone know if a serious attempt at building a German acoustic model has been done by anyone. In searching the archives I've seen one or two similar inquiries, but there were no indications as to whether anything was continued.
Also, just how difficult is it to build up a new acoustic model. I have no experience with what would be required to do this, but I am interested to know how long it might take a computer savi German linguist to do.
many thanks
robert
Hi robert,
1. at least until yesterday there was a thesis offered at http://cortex.informatik.tu-ilmenau.de/~cs/Themen/SJA_Spracherkennung/sja-spracherkenner.html
for the summer semester 2003.
The applicant should make himself aquainted with Sphinx and learn and show how to apply it to German language applications.
2. My understanding of Sphinx is, that one should not try to use it as a replacement for something like ViaVoice, but rather as a very flexible open source tool for building speakerindependent speech controled applications.
For example, I will try to implement it with a whole bunch of language models. One model for each menue and none with more than just some hundred words, which are a mixture of German, English, Latin, and product names. That is to say, I will build one model for each database column and entryfield, since this seems to be the best way to maximize the recognition rate, speakerindependence and speed.
If seems, that with up to 50 words one should train just the words as phonems, while with more words one should try to make a cleverly tailored selection of phonems (in German 'Silben').
A good explantion to start with seems to be man1.html at the SphinxTrain page.
A complete German model only seems to be usefull if one wishes to use Sphinx with rather large vocabularies.
Though the strength of Sphinx seems to be rather the ability to limit the vocabulary to what is really needed, and to switch quickly between language models.
However, I have just studied somewhat the manuals but can not yet play with Sphinx since SphinxTrain does not compile on Suse 8.2's gcc-3.3. Do you have a solution to this?
Regards,
Christoph
Hi,
Thanks for the reply. At the minute I do only need a small vocabulary control application, so Sphinx should do me as it is.
However I still feel that a serious open source replacement for ViaVoice is required. The people over at XVoice seem to think Sphinx is that replacement for the time being. If the feeling is that Sphinx will never scale up to full dictation, then maybe they need to consider another replacement for ViaVoice.
I have not yet made any attempt to use SphinxTrain, I'm still in the stages of asking around to see what has been done. Also I have no training in linguistics, so I'm not really sure of what I would have to do anyway. I might yet come back to use SphinxTrain.
How much actual time effort do you think it would take to build a german acoustic model which would be complex enough for the Turtle example?
yours gratefully
robert