Bird Sound Denoiser
Published:
To synthetically generate the noisy samples, I used a technique called Repet-SIM for separating foreground and background sounds from another noisy bird sound dataset.
According to their GitHub page, "The REPET methods work well when the repeating background has periodically repeating patterns (e.g., jackhammer noise); however, the repeating patterns can also happen intermittently or without a global or local periodicity (e.g., frogs by a pond). REPET-SIM is a generalization of REPET that can also handle non-periodically repeating structures, by using a similarity matrix to identify the repeating elements".
Then these background environmental sounds were added to the clean audio files with a certain SNR. The advantages of creating the noisy audio in this way are as follows:
- First, my primary dataset did not have clean audio targets.
- Second, by mixing in background sounds from an external dataset, I effectively re-created naturally occurring environmental sounds for avian vocalizations.
- The reason I couldn't use Repet-SIM instead of a denoising model is because foreground sounds contained artifacts.
Denoised Audio Samples
Here are a few audio samples that were denoised by DemucsV2. While the model is able to remove various background sounds, it doesn't do well for denoising cricket sounds.| Noisy | Enhanced | Clean |
|---|---|---|