Recordings and the top problems

Removing noise with AI

Recordings and the top problems

It is necessary for me to manually inspect each and every noise recording for voice, cw, etc. before they are included in the training dataset.

Short Recordings

Currently there are many recordings shorter than 30 minutes that are not labeled as tests. I am not training on short recordings, as I assume that a problem was detected, and it was terminated early.

Recordings made for troubleshooting

Sometimes I request short recordings in order for me to troubleshoot issues. These should be as short as possible and labeled appropriately. Please email rmnoise@ournetplace.com if a short recording is intended for me and not training.

Top problems with recordings

  • there should be no voice, cw, wsjt, sine-wave tones, etc.
  • stationary noise should not be recorded.  Stationary noise show as horizontal lines in the noise floor while visualizing the ORIGINAL graph
  • the bandwidth filter should be 2.8KHz for voice and cw
  • the DSP filter shape should rolloff quickly (“sharp” on the ICOM)
  • noise reduction should be off, noise blanker should be off
  • the radio should not filter the low audio frequencies
  • the levels should not change during the recordings
  • Recordings intended for training should be 30 minutes

Here is an example of a bad recording

This recording has two problems.

  • The bandwidth filter appears to be much lower than 2.8KHz
  • There is a horizontal line indicating stationary noise at around 1.8KHz

More on stationary noise

When recordings have stationary noise, the neural net tends to learn that, on average, a good strategy is to simply suppress information at these frequencies. This is likely not ideal for everyone.