I was able to use the simplelm.pl script to create my own language model
[after normalizing the CMU dictionary (i.e. get rid of chars 0,1,2,3)]
At this point sphinx only sometimes figures out what I'm saying,
even when the input was a target sentence used for input to simplelm.pl
Is this mostly because my mic & soundcard differs from whatever
setup the HMM's were trained on? Are there things I can do to improve
the accuracy? (different parms in creating the statistical model?)
Also, are there peices of the planned trainer module that I could help with?
-Jim
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
The audio quality on the channel is important, as well as the volume. On my Compaq Presario 1830 running Linux with the ALSA drivers, I need to set all my levels at about 80%. Anything else and performance suffers; don't forget to turn up the igain (input gain) too. I have found that if i have the igain over about 85% that i get series clipping and performance goes waaay down. With it below 75-80%, the results are much fainter.
I recommend using a headset mic and having the mic audio go through and into the headphones as well, so that you can hear first-hand what the audio quality is like.
Also, it's good for 'natural' speech, but results will degrade with a morce force or over-articulated style -- as when one might get frustrated and try to speak more clearly.
Mic position is important, too. I headset mic with the audio coming into the headphones is good for getting the overall range of levels you need to get the best results from sphinx2, and after that, you can use any of the various mixers to set volumes for you.
Also, some sound cards don't have enough input gain -- particularly some laptops. Perhaps yours is of the kind that takes one of the cheap battery-operated pre-amps...
kevin
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I was able to use the simplelm.pl script to create my own language model
[after normalizing the CMU dictionary (i.e. get rid of chars 0,1,2,3)]
At this point sphinx only sometimes figures out what I'm saying,
even when the input was a target sentence used for input to simplelm.pl
Is this mostly because my mic & soundcard differs from whatever
setup the HMM's were trained on? Are there things I can do to improve
the accuracy? (different parms in creating the statistical model?)
Also, are there peices of the planned trainer module that I could help with?
-Jim
The audio quality on the channel is important, as well as the volume. On my Compaq Presario 1830 running Linux with the ALSA drivers, I need to set all my levels at about 80%. Anything else and performance suffers; don't forget to turn up the igain (input gain) too. I have found that if i have the igain over about 85% that i get series clipping and performance goes waaay down. With it below 75-80%, the results are much fainter.
I recommend using a headset mic and having the mic audio go through and into the headphones as well, so that you can hear first-hand what the audio quality is like.
Also, it's good for 'natural' speech, but results will degrade with a morce force or over-articulated style -- as when one might get frustrated and try to speak more clearly.
Mic position is important, too. I headset mic with the audio coming into the headphones is good for getting the overall range of levels you need to get the best results from sphinx2, and after that, you can use any of the various mixers to set volumes for you.
Also, some sound cards don't have enough input gain -- particularly some laptops. Perhaps yours is of the kind that takes one of the cheap battery-operated pre-amps...
kevin