Adblocker for FM radio
training step:
with labeled audio recordings from sdr tuned to radio station:
- compute features (mel-spectrum? chroma features? NMF?)
- train model on features w/ to learn 'radio' / 'music' / 'noise'
- incorporate prior knowledge:
- radio advertisements usually last 10, 15, 30, or 60 seconds
- advertisements should be identical, so try to come up with features that match across repetitions of advertisement (subsequence DTW?)
operation step:
with input audio stream from sdr tuned to radio station:
- gather observations, using feature extractors
- model current state and transitions using e.g. Hidden Markov Model
- play audio stream, but if ad currently playing, suppress audio (or replace with your own music)