search

drowe67 / freedv-gui

GUI Application for FreeDV – open source digital voice for HF radio
https://freedv.org/
GNU Lesser General Public License v2.1
210 stars 52 forks source link

Dump feature vectors #770

Open drowe67 opened 1 week ago

drowe67 commented 1 week ago

Proposal for a freedv-gui test mode/feature to dump a file of features_out.f32 vectors (36 cols each, x length of sample). These are the vectors at the output of the RADE decoder, that are fed to the FARGAN vocoder for synthesis. By comparing them to a corresponding input set of features_in.f32 vectors, we can objectively measure distortion (using radae/loss.py) of the RADE signal as it passes through freedv-gui.

The features_out.f32 file should start when sync is obtained, and stop when sync stops.

Use cases include:

  1. Testing the implementation of RADE Rx on freedv-gui, ensuring there is no additional distortion introduced into the RADE signal from its path through freedv-gui. In this case we would use a known-good Tx, that takes features_in.f32 and creates a tx wave file that is fed to a freedv-gui Rx.
  2. Testing the implementation of RADE Tx on freedv-gui, by taking a known input speech wavefile, and using it to create a tx wave file. A known good RADE receiver could then be used to test the quality of the Tx waveform, compared to a reference implementation.
  3. OTA optimisation. For example play a known speech source file (which has a corresponding features_in.f32), and measure the loss under various channel and radio settings (ie AGC, Tx drive, or additional filtering).
  4. Test the effect of end user input speech processing, microphones, on RADE performance. If you filter the input, or insert noise suppression, or use a different headset, what is the effect on the RADE loss metric?
tmiw commented 1 week ago

Is this something that would be usable via of the unit test infrastructure being worked on in #761? Or is this intended to be usable during normal usage of FreeDV?

EDIT: I guess for (4) it would have to be usable during regular operation. That brings up the question of how the files get generated (i.e. is a new file created every time PTT is toggled, we keep appending to a single set of files, etc).

drowe67 commented 6 days ago

Initially for unit testing, but I can see some possibilities for OTA tuning/debugging.