Hello,
after I was testing a long time with SimpleText, I now tried SimpleRecord and I noticed that the speech recognition is really bad.
I thought it could be because the dictionary and grammar are not fitting my purpose. So I tried to change the dictionary and the grammar in sphinx (I should do this, right?). However, it did not work as I expected, could you give me a hint on what exactly I have to do?
Until now I tried to modify settings in sphinx-de.xml.
Thank you!
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
The speech recognition is indeed not very good, mostly because the German acoustic models are relatively poor. Could you indicate just how bad speech recognition is?
Some ideas:
- the dictionary does not harm recognition performance, just make sure that all the words in your language model are contained within (watch for warnings when starting up)
- using a good language model is crucial to recognition performance. Unless you want to talk about Pentomino puzzle pieces and how to puzzle elefants out of them, make sure to generate your own. (However, you already did that, right?) Also, performance should somewhat correspond to the size of your language: the smaller the search space, the better your results. In SimpleReco there is the -lm switch which you can use to set a language model via the command line
- there is little I can do about the acoustic models (we're working on it). Make sure to use a headset and a reasonable-quality sound adapter (ideally: USB-headset) in a quiet environment. Casey Kennington has recently integrated Google-ASR as a different source, even though that is relatively cumbersome to set up.
- mostly, the acoustic model is trained on Northern German speech, maybe you're speaking some other dialect?
- finally, do you see performance differences between recognition from a file or from the microphone directly? InproTK uses very low input buffering, which may reduce performance if your computer is not sufficiently fast.
It's good to see you're still working on this!
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hello,
after I was testing a long time with SimpleText, I now tried SimpleRecord and I noticed that the speech recognition is really bad.
I thought it could be because the dictionary and grammar are not fitting my purpose. So I tried to change the dictionary and the grammar in sphinx (I should do this, right?). However, it did not work as I expected, could you give me a hint on what exactly I have to do?
Until now I tried to modify settings in sphinx-de.xml.
Thank you!
The speech recognition is indeed not very good, mostly because the German acoustic models are relatively poor. Could you indicate just how bad speech recognition is?
Some ideas:
- the dictionary does not harm recognition performance, just make sure that all the words in your language model are contained within (watch for warnings when starting up)
- using a good language model is crucial to recognition performance. Unless you want to talk about Pentomino puzzle pieces and how to puzzle elefants out of them, make sure to generate your own. (However, you already did that, right?) Also, performance should somewhat correspond to the size of your language: the smaller the search space, the better your results. In SimpleReco there is the -lm switch which you can use to set a language model via the command line
- there is little I can do about the acoustic models (we're working on it). Make sure to use a headset and a reasonable-quality sound adapter (ideally: USB-headset) in a quiet environment. Casey Kennington has recently integrated Google-ASR as a different source, even though that is relatively cumbersome to set up.
- mostly, the acoustic model is trained on Northern German speech, maybe you're speaking some other dialect?
- finally, do you see performance differences between recognition from a file or from the microphone directly? InproTK uses very low input buffering, which may reduce performance if your computer is not sufficiently fast.
It's good to see you're still working on this!