I've recored a small audio dataset (about 4 min), containing only 4 words said in different order. I know this is far away from enough to work, but I expected it to at least recognize (and transcript) the original audio files (in a batch mode) but it didn't, is this normal.
recording a small dataset followed by creating an acoustic model from it then testing that model with the original data is the most basic test I can think of before recording more data, so I really need to know if I'm doing things right.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I did something similar when I was starting. I don't remember exactly how much data I used, but it was comparable. Sounds to me like something is going wrong, and you ought to figure that out before spending a lot of effort gathering more data.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
qetoom, i have the same problem that you. i try to test with my own training sound files and in batch mode recognizer the result is 0%, and any hypotesis.
My model.props is right, how to resolve this problem you?
Thanks, Rubén.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
As we have tried to say in this forum again and again, that amount of training data will only give you bad acoustic model training. With bad acoustic model, it is not a wonder that decoder will give poor result.
From my experience we train small data (20 word only single speaker the over all duration is 6 min. )
The likelyhood is 30.124 )
When I run batch test I get 100% accuracy
When running live test I get no results
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I've recored a small audio dataset (about 4 min), containing only 4 words said in different order. I know this is far away from enough to work, but I expected it to at least recognize (and transcript) the original audio files (in a batch mode) but it didn't, is this normal.
recording a small dataset followed by creating an acoustic model from it then testing that model with the original data is the most basic test I can think of before recording more data, so I really need to know if I'm doing things right.
I did something similar when I was starting. I don't remember exactly how much data I used, but it was comparable. Sounds to me like something is going wrong, and you ought to figure that out before spending a lot of effort gathering more data.
qetoom, i have the same problem that you. i try to test with my own training sound files and in batch mode recognizer the result is 0%, and any hypotesis.
My model.props is right, how to resolve this problem you?
Thanks, Rubén.
it appears that I've miss configured the model.props
file, now every thing works fine.
As we have tried to say in this forum again and again, that amount of training data will only give you bad acoustic model training. With bad acoustic model, it is not a wonder that decoder will give poor result.
Please take a look of what I wrote about half a year ago,
http://www.cs.cmu.edu/~archan/10CommonPitfalls_ST.html
From my experience we train small data (20 word only single speaker the over all duration is 6 min. )
The likelyhood is 30.124 )
When I run batch test I get 100% accuracy
When running live test I get no results