Hi guys, we are seeing amazing scores already, a product of the hard work put into the just concluded hackathon🙇. There is still enough time to learn and work,💪 so I won't be sharing my notebook for the hackathon. Nevertheless, I'll share my findings, and tips on how you can beat my current score on the leaderboard.
0) Data is about 99.9% clean, for me I only found one, so there may be no need to look for wrongly labeled audio files.
1) Do good augmentations when converting to spectrograms, for me "Removing silence worked well"
2) You'll definitely want to ensemble with an ASR. The hugging face basic tutorial was all I needed for the hack. You can find the official tutorial here:
3) Ensembling diverse approaches worked better than ensembling the same model
4) FastAI's FastAudio approach has excellent voice configs used to generate spectrograms, which you may want to explore.
5) If you have no idea how to start, use the starter notebook. Nevertheless, I made a comprehensive tutorial weeks ago here for noise audio classification:
https://github.com/osinkolu/DataFest-Africa-Noise-Pollution-Classification-Challenge "if you find this repo insightful, don't forget to leave a star" 🌟
Expecting to see y'all on the leaderboard.....😉