Other languages

[Mireia Farrús]

Hello,

We need to automatically label prosody for other languages and accents: British English, French, Spanish, etc.
I would like to know if there is an easy way to adapt AuToBI to these languages. What should be changed with respect to the current AE annotation?

Thank you very much,

Mireia

[Andrew Rosenberg]

Hi Mireia,

Thanks for your interest here.  

If you have annotated data for the language you're working on, you can use the training routines to retrain models specific to those languages.   There might be some minor modifications necessary to the input file reading routines to appropriately read a different inventory of labels.  If this is needed, let me know and I can work through the details with you.

You could use the American English models and apply them to another language.  This is not quite ideal, but we've done some experiments on cross-language prediction of phrase boundaries and pitch accents. The results aren't as horrible as you might suspect.  (http://www.cs.columbia.edu/speech/PaperFiles/2013/soto-Columbia-icassp13-v2.pdf, http://speech.cs.qc.cuny.edu/pubs/rosenberg2012cross.pdf). I wouldn't trust the type classification very much though.

If you only have unlabeled data the situation is a little bit more difficult.  I haven't implemented any unsupervised adaptation in AuToBI. This is something I've been thinking about.  It will take some experimentation to figure out the best way to do this. It's an important direction for future work but not likely to come online in the short term.

-Andrew

--
You received this message because you are subscribed to the Google Groups "AuToBI Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to autobi-users...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.

[Mireia Farrús Cabeceran]

Thank you very much for your quick response, Andrew.

I was aware of your Cross-Language Prominence and Phrase Boundary Detection; I think you should first try to retrain the models using the training routines, as you state in the first option.

I will let you know how it proceeds.

Best regards,

Mireia


El dijous 27 de febrer de 2014 14:48:14 UTC+1, Andrew Rosenberg va escriure: