I'm now officially crazy, having tried to several days to diagnose why one particular waveform isn't recognized correctly. I just discovered (quite by accident) that the result depends on the amplitude of the waveform. My original waveform, which is quite "loud", is recognized incorrectly. A modified waveform (scaled down by 5dB) is recognized correctly.
My first reaction is that there must be an arithmetic overflow somewhere. Does anyone have suggestions for what might be going on or where I should begin looking?
Carl Hagenmaier
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I'm interested in hearing the response to this. In theory, the Fourier Transform (the first step in feature extraction) should be impervious to amplitude.
If you are willing/able to do batch mode recognition, you can try cepstral mean normalization, but this is aimed at compensating for channel distortion.
Hopefully someone has an answer...
Robbie
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I'm now officially crazy, having tried to several days to diagnose why one particular waveform isn't recognized correctly. I just discovered (quite by accident) that the result depends on the amplitude of the waveform. My original waveform, which is quite "loud", is recognized incorrectly. A modified waveform (scaled down by 5dB) is recognized correctly.
My first reaction is that there must be an arithmetic overflow somewhere. Does anyone have suggestions for what might be going on or where I should begin looking?
Carl Hagenmaier
I'm interested in hearing the response to this. In theory, the Fourier Transform (the first step in feature extraction) should be impervious to amplitude.
If you are willing/able to do batch mode recognition, you can try cepstral mean normalization, but this is aimed at compensating for channel distortion.
Hopefully someone has an answer...
Robbie