How to handle amplitude/gain differences between training sounds and live microphone sounds

[Toine db]

I'm looking into al possible ways how to handle differences in amplitude, as a newbie I call it sound volume.

My training recordings had a amplitude of 7000-8000 and my micrphone recorded 1000-2000.
Surely to large a difference to use for accurate recognition.

I got the following comment in a very old thread

Ideally you need a good recording level normalizer, probably we need to improve AGC or implement very quick CMN.

.
So

• Is it better to train with multiple different redording levels
  or to use a AGC to control the microphone input ?

• Anyone recommend and AGC or normalizer ?

I'm relative new in the world of sound and its terms, so I hope someone can help me to remove this problem.

[Toine db]

anyone?

[Nickolay V. Shmyrev]

Is it better to train with multiple different redording levels or to use a AGC to control the microphone input ?

Mathematically it is better to have good AGC than to train on multiple levels. Another alternative would be to use a classifier which can learn from more complex unnormalized data (deep neural networks for example are way better classifiers), but this goes beyond current capabilities of CMUSphinx toolkit.

Anyone recommend and AGC or normalizer ?

You can ask in google

http://dsp.stackexchange.com/questions/3403/automatic-gain-control-for-voice

[Toine db]

Thanks for the info Nickolay.

And of course Google knows it :-)