Different volume levels are leading to highly differently recognition rates

[Frank Sippel]

Hi!

I am currently evaluating the performance of Pocketsphinx using Matlab and the Python wrapper for Pocketsphinx (https://github.com/bambocher/pocketsphinx-python). I'm currently stuck in a weird situation, because I first normalize to the maximum absolute value and then applying different scale values. In Matlab Code :

result=result ./ max(abs(result(:)));
result=result * normalizingfactor;

"result" is the audio signal.
Then I send it to Pocketsphinx and afterwards it gets evaluated by the nist scoring toolkit.
So here are the results :
Correct (x-Axis -> normalizing factor, y-Axis -> percentages):

Total Error (x-Axis -> normalizing factor, y-Axis -> percentages):

Do you have any suggestions how to overcome this problem and get a flat curve (disregarding quantization issues)?
Or at least finding the best value?

Thanks in advance!
~Frank

[Nickolay V. Shmyrev]

Recognizer adapts to any audio level but does that only after the first utterance so the test you created is not very meaningful I suppose, you just do not leave the recognizer the time to adapt.
True decoding accuracy does not depend on volume level.

You can read this forum about cepstral mean normalization.

[Frank Sippel]

Thank you very much! By putting all test sound files together into one large sound file, I'm getting the same results except (like you already said) for the first utterance :-)