{
  "id": 160759,
  "title": "Lessons learned from 2019 Freesound audio tagging competition",
  "url": "/competitions/birdsong-recognition/discussion/160759",
  "author_name": "PAB97",
  "post_date": "2020-06-22T14:49:04.703000",
  "votes": 21,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hello everyone, </p>\n\n<p>I'm just starting this competition. I remember doing one year ago almost exactly day to day the Freesound audio tagging classification. These competitions are a bit special since they involve audio and you need an extra \"step\" which is converting audio files to images.</p>\n\n<p>You can find a notebook that I wrote one year ago on how to convert audio to images: \n<a href=\"https://www.kaggle.com/rftexas/converting-sounds-into-images-a-general-guide\">Converting sounds into images: a general guide</a></p>\n\n<p>Moreover, a learnt some useful techniques along the way to succeed in audio competitions:</p>\n\n<ul>\n<li><p>Usually, Mixup augmentation techniques work quite well for melspectrogram data. I remember that it gave a nice boost on FAT 2019 competition.</p></li>\n<li><p>TTA worked well as well. I remembered I used to use fast.ai library where they had a nicely-implemented TTA callback. It is quite easy to use TTA with it.</p></li>\n<li><p>Try changing the sampling rates and other parameters you choose when converting an audio to an image. Best teams tried multiple sampling rates to create blends.</p></li>\n</ul>\n\n<p>I hope this helps and might enrich this thread with more lessons if I remember something additional ;)</p>",
  "messages": [
    {
      "id": 896976,
      "postDate": "2020-06-22T14:49:04.703Z",
      "content": "<p>Hello everyone, </p>\n\n<p>I'm just starting this competition. I remember doing one year ago almost exactly day to day the Freesound audio tagging classification. These competitions are a bit special since they involve audio and you need an extra \"step\" which is converting audio files to images.</p>\n\n<p>You can find a notebook that I wrote one year ago on how to convert audio to images: \n<a href=\"https://www.kaggle.com/rftexas/converting-sounds-into-images-a-general-guide\">Converting sounds into images: a general guide</a></p>\n\n<p>Moreover, a learnt some useful techniques along the way to succeed in audio competitions:</p>\n\n<ul>\n<li><p>Usually, Mixup augmentation techniques work quite well for melspectrogram data. I remember that it gave a nice boost on FAT 2019 competition.</p></li>\n<li><p>TTA worked well as well. I remembered I used to use fast.ai library where they had a nicely-implemented TTA callback. It is quite easy to use TTA with it.</p></li>\n<li><p>Try changing the sampling rates and other parameters you choose when converting an audio to an image. Best teams tried multiple sampling rates to create blends.</p></li>\n</ul>\n\n<p>I hope this helps and might enrich this thread with more lessons if I remember something additional ;)</p>",
      "rawMarkdown": "Hello everyone, \n\nI'm just starting this competition. I remember doing one year ago almost exactly day to day the Freesound audio tagging classification. These competitions are a bit special since they involve audio and you need an extra \"step\" which is converting audio files to images.\n\nYou can find a notebook that I wrote one year ago on how to convert audio to images: \n[Converting sounds into images: a general guide](https://www.kaggle.com/rftexas/converting-sounds-into-images-a-general-guide)\n\nMoreover, a learnt some useful techniques along the way to succeed in audio competitions:\n\n- Usually, Mixup augmentation techniques work quite well for melspectrogram data. I remember that it gave a nice boost on FAT 2019 competition.\n\n- TTA worked well as well. I remembered I used to use fast.ai library where they had a nicely-implemented TTA callback. It is quite easy to use TTA with it.\n\n- Try changing the sampling rates and other parameters you choose when converting an audio to an image. Best teams tried multiple sampling rates to create blends.\n\nI hope this helps and might enrich this thread with more lessons if I remember something additional ;)",
      "votes": 21
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "896976": "Hello everyone, \n\nI'm just starting this competition. I remember doing one year ago almost exactly day to day the Freesound audio tagging classification. These competitions are a bit special since they involve audio and you need an extra \"step\" which is converting audio files to images.\n\nYou can find a notebook that I wrote one year ago on how to convert audio to images: \n[Converting sounds into images: a general guide](https://www.kaggle.com/rftexas/converting-sounds-into-images-a-general-guide)\n\nMoreover, a learnt some useful techniques along the way to succeed in audio competitions:\n\n- Usually, Mixup augmentation techniques work quite well for melspectrogram data. I remember that it gave a nice boost on FAT 2019 competition.\n\n- TTA worked well as well. I remembered I used to use fast.ai library where they had a nicely-implemented TTA callback. It is quite easy to use TTA with it.\n\n- Try changing the sampling rates and other parameters you choose when converting an audio to an image. Best teams tried multiple sampling rates to create blends.\n\nI hope this helps and might enrich this thread with more lessons if I remember something additional ;)"
  }
}