Menu

Creating acoustic model in 48 khz

Help
2014-08-29
2014-08-29
  • Colin Gilbert

    Colin Gilbert - 2014-08-29

    Hi,

    I am making an acoustic model for small vocabulary command & control and wish to increase the accuracy as much as possible. The manual says to only use 8000 and 16000hz, but portability and performance are not concerns: My computer has a stationary receiver + noise cancelling headset which does not travel and is powered by a 3.4ghz Xeon + 32gigs RAM. Also, it would be a hassle to resample input audio, as it means yet more coding/system configuration (my audio card does not go below 44100 :( ) and probable loss of signal quality. Also, I wish to limit the data to what will be actually used and a dataset trained on the entire English language may make more mistakes than a small, specialized one for each app.

    My idea is to use a lowpass of 500 and a highpass of 20000hz. Are there any pitfalls which I may be unaware of and other settings that should be modified?

    Thank you.
    Colin

     

    Last edit: Colin Gilbert 2014-08-29
    • Nickolay V. Shmyrev

      Hello Colin

      You can use 20000hz range, you need to care about properly setting up the feature extraction parameters. However, I don't think it will bring you much accuracy over 16khz.

       

Log in to post a comment.

MongoDB Logo MongoDB