VoxForge
From the HTK mailing list archives:
> Hi everyone,
>
> I am trying to implement a continuous speech recognition in Spanish
> Language. I followed the indications to make the tri-gram language
> model presented in the HTK book. I am using
the HDecode tool. My
> firsts results are very poor (WER >= 35%). So I tuned various
> parameters in HDecode, but know I need to make tuning of the language
> model parameter. Any user can help me in that? for example, what
> parameter in the to language model generation are recommended to
> continuous speech?
Did you try to follow htk wsj1 reciept? It has almost everything
required I think:
http://www.inference.phy.cam.ac.uk/kv227/htk/
http://www.inference.phy.cam.ac.uk/kv227/lm_giga/
with all beams used and lm factors. Though for really large vocabulary lm factor should be smaller (around 6-8).
A subdirectory or file hmmlog13 already exists.
exe\HHEd.exe -A -D -T 1 -H hmmlog12/macros -H hmmlog12/hmmdefs -M hmmlog13 tree.hed triphoneslog1
No HTK Configuration Parameters Set
HHEd
59/59 Models Loaded [5 states max, 1 mixes max]
RO 100.00 ''
Setting outlier threshold for clustering
RO->LS stats
and loading state occupation stats
Stats loaded for 59 models
Mean Occupation Count = 14.841730
TR 0
Adjusting trace level
WARNING [-2631] QuestionCommand: No items for question R_Silence
in exe\HHEd.exe
WARNING [-2631] QuestionCommand: No items for question R_Nasal
in exe\HHEd.exe
*
^
Error { expected
ERROR [+7230] EdError: item list parse error
FATAL ERROR - Terminating program exe\HHEd.exe