Reducing Noise in Feynman Lectures

Question

There are quite a lot of old Feynman lecture recordings on YouTube, that unfortunately suffer from very noisy audio. I'm not a sound-design professional, hobbyist; in fact I'm quite possibly tone deaf. I am however an engineer, so I had to give it a whack.

I tried applying a couple of filters in Audacity on a sample lecture

• Automatic noise reduction filter,
• Low-pass filter for frequencies above 2 kHZ.
• High-pass filter for frequencies below 300 Hz

I was actually quite pleased with the results. Compare the two samples - original YouTube video and the filtered result. I would say that the voices in the original somehow sound more real, so there's room for improvement.

I'm looking for input and suggestions (filters, settings) on how to improve this further.

Answer 1

Unless there is a pattern to the audio, the rule is generally crap in, crap out with audio. You can remove the parts of the spectrum that the noise is primarily on or try to do pattern matching to identify what is the noise and what is signal, but separating audio streams exactly is impossible once things are mixed together.

This is why you end up with problems with the sound of the voice becoming unnatural with you low and high pass filters. At those levels, they are significantly cutting off the frequencies of the voice itself, most likely both the tone of the voice and the part that we hear syllables from.

If there is any hope of improving it, you would want to try using a gate to remove noise when someone isn't talking and possibly try a pattern matching filter, but it still likely isn't going to sound like a clean recording, it will just be a slightly easier to understand noisy recording most likely.

Answer 2

This is pretty much all you can do in audiacity. For such jobs you need a audio restauration software. There you can differentiate noise from Tones in the frequency Spektrum. Then by modulating those you can get good Results.