This project collects multiple synchronized audio files or streams of the same "event" [in particular, a lecture] and filters out idiosyncrasies in each recording like silences, pops and bumps, and background noise (like talking) to produce one, robust recording.
Currently, I have four artificial recordings for a sanity check - verifying the program reconstructs the proper signal if it has very little noise in the input. When the program filtering_demo.py is run, it takes the three modified recordings and reconstructs the original. Note that the records are already synchronized perfectly and, in many places, are the same exact waveform.
darkness.wavis the base; it is the beginning of Darkness by LEVV.darkness_click.wavincludes a pen-click sound at the beginning.darkness_silence.wavfades out and back in near the end.darkness_speak.wavincludes some spoken text in the middle, in particular, "One" by La Dispute.darkness_filtered.wavis the reconstructed sound, after the script is run.
I manually synchronized three recordings from our phones my friends and I made one night, then ran the filtering algorithm. There was significant distortion in the output, but the best method I've found so far is the combination of a "complex median" and feathering [linear blending / cross-fading] the sections.
In the ClassCapture_Synth/synth directory, run:
python filtering_demo.py
and the filtered WAV files will appear.