How can I filter a noise with pattern type in audio?

Question

I have a set of audio files rescued from a source of phonograph record like a gramophone. Very old family stuff records from my elders back in year 1946. A relative has digitalized these with a turntable and gave the records to me in lossless WAV Format. The audios are just conversations of my relatives in cotidiane situations. So it is just vocals materials but mixed with a heavy quantity of noise. Noise is mostly concentrated in the mid-high frequencies, like 1Khz-10Khz.

So the records are VERY noisy. SNR is like 0dB or worst. My mission is to filter out the noise.

One advantage is that noise exhibits a cyclical pattern. I tried some classic vinyl restoration stuff like decracking o declicking with Sound Forge, DART 24 Pro. Results were unsuccessful. The best result I had is the "Noise reduction" effect in Audacity, which did output relatively good. But this last one just filters the frequencies where the noise is with narrow FFT. So the vocals content loses the high-mid frequencies. Leaving only mid or mid-bass content. The noise got down but voices are unintelligible.

So I think perhaps I could turn to an Artificial Intelligence solution?

But I have no clue from where can I start. I have some understanding of theory of AI, the Perceptron, neural networks, Back-propagation, but I never applied them to audio. Perhaps a Convolution Neural network could be a solution, but how could I train it?. If helps, I have programming skills in C# and knowledge about digital linear filters like IIR and FIR. But equalization based solutions are not, what I need.

Below I offer an example track for your consideration.

Audio track for example

What can I do for filter this kind of cyclical pattern noise?

Answer 1

Simple idea: let some pros do the work. Send the audio through a cellphone to another one via a WhatsApp, Duo or whatever voice call app, even classic network call works, dealer's choice. The noise reduction algorithms in cellphones are designed to do exactly what you want to do. Reduce noise while conserving human speech. The outcome will depend to some degree on the market, the phone's firmware was designed for. US customers like it, when the background noise is cut out completely, accepting some degradation of speech. European customers prefer more naturally sounding outcomes with some noise left.

I think this would be worth a shot.

Answer 2

This is actually an excellent place to start with ML/DNN tools. Noise Reduction, Speech Processing and Recognition are driving a lot of the innovation in sound in this space.

Recurrent Neural Networks and LSTM models are good at identifying patterns - which can be useful in this context. https://jmvalin.ca/demo/rnnoise/

If you’ve got an NVIDIA GPU you could also try experimenting with the RTX noise suppression that was quite popular a few years ago - but it was designed for live noise suppression for broadcasting. Recently was hacked to allow more cards:

https://arstechnica.com/gaming/2021/04/nvidia-adds-official-rtx-voice-noise-cancellation-to-non-rtx-cards/

You’ll never be able to clean up broadband noise this severe completely - but it’s a perfect reason to explore some of these tools and concepts.