search

mmcauliffe / Conch-sounds

This package contains functions for converting wav files into auditory representations and comparing them
MIT License
54 stars 11 forks source link

Using this as an Audio Quality tool? #6

Closed migueljette closed 7 years ago

migueljette commented 7 years ago

Hi Michael,

Do you think there would be a way to use your tool here for some sort of Audio Quality assessment tool?

I'm looking for a tool that could give me an idea of how "clean" an audio is. Ultimately, I need this tool to have an idea of how hard it will be to transcribe the file. My thought was that I could use your tool here to compare the audio to N audios representing N potential quality levels, and choose the quality closest to my audio. For example, N=3.

If I compare my audio and it most closely matches the "cafe noise" audio, then I can assume the audio was recorded in a cafe (or an environment close to that). Does that make sense? What do you think?

Thanks for your input! Again, nice looking tool!

Cheers, Miguel

mmcauliffe commented 7 years ago

So I don't know that the tools here would be particularly helpful for you in determining the type of noise in a given file. Most of the work here is focused on clean, lab-recorded audio.

What you'd probably want is something like a voice-activity detector coupled with a noise estimator (for places where no voice was detected). I think it'd be somewhat straightforward to train a neural network on clean data with superimposed cafe/driving/other noise as a voice activity detector (i.e., for a given frame of audio, is there someone speaking).

You could run that on the file, create a long-term average spectrum of the frames without voice activity to create a noise profile that you can match to known noise types. You could probably use the MFCC/sprectral estimation in acousticsim for that last step.

Hope that helps!

migueljette commented 7 years ago

Hi Michael,

Thanks for your response! It might come in handy in the future!