Here are the scripts I used to train an HTK recognizer using the spanish data I found on Voxforge (approx. 20 hours). The data are not perfect but one can build a baseline recognizer using them. The recipe is based on Keith Vertanen's scripts for HTK TIMIT + WSJ ASR training (http://www.keithv.com/software/htk). Experimenting with Makefiles these days, I wrote the entire process as a Makefile and I have to say it can be quite convenient for this task. However, there are many things that I haven't really fixed and haven't yet fully harnessed the power of make :) This is yet to be done. The idea is to just write (after updating the paths in the makefile): make all and wait. Of course you may need to delve deeper into the code if you have more specific needs. I haven't really tested it yet in an independent setup so it may need minor modifications to work for a different dataset or on a different machine. I will be working on it and this code should only be considered as a working version and not an official release. In case you just need the spanish acoustic models for HTK, please feel free to contact me and I could share them with you. If you have spanish data you would like to train an ASR on, we could work on that together. Nassos Katsamanis
thomaskisler/spanish_voxforge_htk_asr
Makefile recipe to train htk acoustic models using the spanish data on voxforge. It should be working by just updating the paths and typing: make all.
Perl