FM Radio Super Resolution
This is a model designed to upscale recordings of FM radio. It was created using the Aero architecture with a modified version of the "Aero" project:
Usage:
- Check out and install the modified version of the Aero project and put the checkpoint file in the working directory.
- If needed, convert the audio recordinng to 44.1 khz stereo.
- Call the new project with:
python predict.py experiment=aero_441-441 +filename="PATH_TO_AUDIO_FILE" +output="PATH_TO_OUTPUT_DIRECTORY" checkpoint_file="PATH_TO_CHECKPOINT_FILE"
The model should be able to:
- Reduce stereo hiss
- Reduce certain kinds of static/interference
- Reduce 19khz stereo pilot
- Reduce overly warm bass frequencies (boost them 3 or so dB if you miss them)
- Restore lost high & low frequencies (generally only up to 16khz since most training data is MP3-based)
Things to be aware of:
- The model isn't currently great at removing static from the mono (L+R) channel, as those have been dificult to reproduce using the equipment I have.
- The "predict" script will normalize the audio after running inference on the whole file, so it likely will be at a different volume than the source track.
- The model was trained from audio digitized directly off the radios used, but should work decently for recordings digitized from cassette tapes, etc.
Transmitters Used:
- Scosche "TuneIn" FM transmitter
- nulaxy KM28 FM transmitter
- Monster iPod/iPhone FM transmitter
Radios Used:
- JVC PC-RM100 System (circa 1984)
- Sony Walkman WM-F76 (Circa 1987)
- Sony CFS 212 Boombox (circa 1988)
- Aiwa CX NA202 System (circa 1999)
- Kaito KA1103 (originally releaed 2004)
- Eton Mini (circa 2017)
- Retroid Pocket Flip (Android device with FM tuner-2023)
Tracks Used:
Version History:
- 1.0.0 (7/22/2024) First version published to Hugging Face