Hi All!
\n\nWhile playing back a few audio files, I've realized, there is a huge amount of variance in the sound levels and there is some background wind/external noise.
\n\nI'll keep updating this post as I find more techniques for pre-processing the audio, that might be helpful:
\n\nGood Luck!
", "messages": [ { "id": "887591", "postDate": "06/15/2020 18:52:41", "content": "Hi All!
\n\nWhile playing back a few audio files, I've realized, there is a huge amount of variance in the sound levels and there is some background wind/external noise.
\n\nI'll keep updating this post as I find more techniques for pre-processing the audio, that might be helpful:
\n\nGood Luck!
", "rawMarkdown": "Hi All! \n\nWhile playing back a few audio files, I've realized, there is a huge amount of variance in the sound levels and there is some background wind/external noise. \n\nI'll keep updating this post as I find more techniques for pre-processing the audio, that might be helpful:\n\n- SpecMix: A Simple Data Augmentation and Warm-up Pipeline to Leverage Clean and Noisy Set for Efficient Audio Tagging by @ebouteillon [Paper](http://dcase.community/documents/challenge2019/technical_reports/DCASE2019_Bouteillon_27_t2.pdf), [Code](https://github.com/ebouteillon/freesound-audio-tagging-2019)\n- A few data augmentation techniques using fastai audio [Github](https://github.com/zcaceres/fastai-audio/blob/master/DataAugmentation.ipynb)\n- [Talk on audio classification](https://www.youtube.com/watch?v=QEEBNF0aeeg)\n- [Librosa package for music and audio pre-processing](https://librosa.github.io/librosa/index.html)\n\nGood Luck!", "votes": null }, { "id": "888037", "postDate": "06/16/2020 04:39:53", "content": "How about Fourier transforms?
", "rawMarkdown": "How about Fourier transforms?", "votes": null }, { "id": "890786", "postDate": "06/17/2020 17:42:33", "content": "The main use-case / purpose of this competition is modelling for the domain mismatch between training and test data.\nThe train-audio files are clean recordings, while the test set contains noisy recordings made in the wild. I think using augmentation techniques that introduce background noise in the training data, as well as overlapping vocalizations, are going to be very important.
", "rawMarkdown": "The main use-case / purpose of this competition is modelling for the domain mismatch between training and test data.\nThe train-audio files are clean recordings, while the test set contains noisy recordings made in the wild. I think using augmentation techniques that introduce background noise in the training data, as well as overlapping vocalizations, are going to be very important.", "votes": null }, { "id": "893987", "postDate": "06/20/2020 05:21:14", "content": "nice info
", "rawMarkdown": "nice info", "votes": null }, { "id": "954353", "postDate": "08/01/2020 16:49:24", "content": "MFCCs could also be an approach I feel
", "rawMarkdown": "MFCCs could also be an approach I feel", "votes": null }, { "id": "968521", "postDate": "08/13/2020 04:50:27", "content": "Very helpful.
\nš
Thanks for this :)
", "rawMarkdown": "Thanks for this :)", "votes": null } ], "comments": [ { "id": 888037, "author_name": "nxrprime", "author_url": "", "post_date": "06/16/2020 04:39:53", "content": "How about Fourier transforms?
", "votes": null, "replies": [] }, { "id": 890786, "author_name": "dhruvrnaik", "author_url": "", "post_date": "06/17/2020 17:42:33", "content": "The main use-case / purpose of this competition is modelling for the domain mismatch between training and test data.\nThe train-audio files are clean recordings, while the test set contains noisy recordings made in the wild. I think using augmentation techniques that introduce background noise in the training data, as well as overlapping vocalizations, are going to be very important.
", "votes": null, "replies": [ { "id": 968521, "author_name": "palaksood97", "author_url": "", "post_date": "08/13/2020 04:50:27", "content": "Very helpful.
\nš
nice info
", "votes": null, "replies": [] }, { "id": 954353, "author_name": "digvijayyadav", "author_url": "", "post_date": "08/01/2020 16:49:24", "content": "MFCCs could also be an approach I feel
", "votes": null, "replies": [] }, { "id": 971038, "author_name": "raoofnaushad", "author_url": "", "post_date": "08/15/2020 05:55:59", "content": "Thanks for this :)
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "887591": "Hi All! \n\nWhile playing back a few audio files, I've realized, there is a huge amount of variance in the sound levels and there is some background wind/external noise. \n\nI'll keep updating this post as I find more techniques for pre-processing the audio, that might be helpful:\n\n- SpecMix: A Simple Data Augmentation and Warm-up Pipeline to Leverage Clean and Noisy Set for Efficient Audio Tagging by @ebouteillon [Paper](http://dcase.community/documents/challenge2019/technical_reports/DCASE2019_Bouteillon_27_t2.pdf), [Code](https://github.com/ebouteillon/freesound-audio-tagging-2019)\n- A few data augmentation techniques using fastai audio [Github](https://github.com/zcaceres/fastai-audio/blob/master/DataAugmentation.ipynb)\n- [Talk on audio classification](https://www.youtube.com/watch?v=QEEBNF0aeeg)\n- [Librosa package for music and audio pre-processing](https://librosa.github.io/librosa/index.html)\n\nGood Luck!", "888037": "How about Fourier transforms?", "890786": "The main use-case / purpose of this competition is modelling for the domain mismatch between training and test data.\nThe train-audio files are clean recordings, while the test set contains noisy recordings made in the wild. I think using augmentation techniques that introduce background noise in the training data, as well as overlapping vocalizations, are going to be very important.", "893987": "nice info", "954353": "MFCCs could also be an approach I feel", "968521": "Very helpful. \nš", "971038": "Thanks for this :)" }, "source": "meta" }