Menu

Effect of training data amplitude on accuracy

dovark
2013-05-24
2014-07-04
  • dovark

    dovark - 2013-05-24

    Hi,

    If I have clean training data but with low amplitude ( average speech amplitude <= 0.02) will the models work well for test data with twice or more average speech amplitude?

    Here are the amplitude distribution plots for two wav files.
    https://docs.google.com/file/d/0BzNdoGke8tyRQmY1M2MxRnl2WjA/edit?usp=sharing
    Above file shows amplitude distribution for an audio file from the training data (WSJ1 database). Maximum amplitude is ~ 0.05. Average speech amplitude is <= 0.02 (I didn't calculate it by removing silence frames, just a guess from the figure)

    Below image shows amplitude histogram for my test file. Max amplitude is 0.08, Average speech amplitude is >= 0.04.
    https://docs.google.com/file/d/0BzNdoGke8tyReE1acGpMWFM4c1U/edit?usp=sharing

    If the training data amplitude is so low, can I just amplify them by X dB and retrain the models?

     
  • dovark

    dovark - 2013-05-24

    Regarding plots, Y axis is frequency, X axis is magnitude of amplitude.

     
  • Nickolay V. Shmyrev

    If I have clean training data but with low amplitude ( average speech amplitude <= 0.02) will the models work well for test data with twice or more average speech amplitude?

    Amplitude (first cepstrum coefficient) is actually normalized during CMN. So the accuracy will be the same unless you have to high amplitude due to clipping or two low amplitude due to precision loss.

     
  • Serotonergic

    Serotonergic - 2014-07-04

    Thanks Nickolay. This is helpful for my case as well. Is there an actual value you would recommend for the too "high" and too "low" situation? What in your experience would you quantify these as?

     

Log in to post a comment.

Want the latest updates on software, tech news, and AI?
Get latest updates about software, tech news, and AI from SourceForge directly in your inbox once a month.