{
  "id": 448504,
  "title": "My Fails (Solution)",
  "url": "/competitions/bengaliai-speech/writeups/jainam213-my-fails-solution",
  "author_name": "",
  "post_date": "2023-10-19T22:33:54.312417600Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>1)I'm surprised none of the high scoring teams converted audio to spectrograms and applied image algorithms on them. Spent quite some time trying this but nothing worked.</p>\n<p>2)Next I tried running noise reduction on the audios during infrence. Doesn't add a lot of overhead and noise reduction algorithms work pretty well, this was just making my score worse though. I still think this is a viable option and there should be some(even though slight) improvement especially in OOD domains if we find the sweet spot for noise reduction.</p>\n<p>3)What i tried next was implementing/using a linear probe instead of finetuning. So finetuning effects OOD generalization and theoretically linear probes should preform better. (My half baked implementation couldn't though) . I was out of time at this point.</p>\n<p>Hopefully this helps someone somewhere.</p>",
  "messages": [
    {
      "id": "2489327",
      "postDate": "10/19/2023 22:33:54",
      "content": "<p>1)I'm surprised none of the high scoring teams converted audio to spectrograms and applied image algorithms on them. Spent quite some time trying this but nothing worked.</p>\n<p>2)Next I tried running noise reduction on the audios during infrence. Doesn't add a lot of overhead and noise reduction algorithms work pretty well, this was just making my score worse though. I still think this is a viable option and there should be some(even though slight) improvement especially in OOD domains if we find the sweet spot for noise reduction.</p>\n<p>3)What i tried next was implementing/using a linear probe instead of finetuning. So finetuning effects OOD generalization and theoretically linear probes should preform better. (My half baked implementation couldn't though) . I was out of time at this point.</p>\n<p>Hopefully this helps someone somewhere.</p>",
      "rawMarkdown": "1)I'm surprised none of the high scoring teams converted audio to spectrograms and applied image algorithms on them. Spent quite some time trying this but nothing worked.\n\n2)Next I tried running noise reduction on the audios during infrence. Doesn't add a lot of overhead and noise reduction algorithms work pretty well, this was just making my score worse though. I still think this is a viable option and there should be some(even though slight) improvement especially in OOD domains if we find the sweet spot for noise reduction.\n\n3)What i tried next was implementing/using a linear probe instead of finetuning. So finetuning effects OOD generalization and theoretically linear probes should preform better. (My half baked implementation couldn't though) . I was out of time at this point.\n\nHopefully this helps someone somewhere.",
      "votes": null
    },
    {
      "id": "2491339",
      "postDate": "10/21/2023 14:32:39",
      "content": "<p>Interesting approaches! The idea of converting audio to spectrograms is particularly intriguing. Thanks for sharing your experiences – they will surely help others in future projects. Best of luck in upcoming contests!</p>",
      "rawMarkdown": "Interesting approaches! The idea of converting audio to spectrograms is particularly intriguing. Thanks for sharing your experiences – they will surely help others in future projects. Best of luck in upcoming contests!",
      "votes": null
    },
    {
      "id": "2492031",
      "postDate": "10/22/2023 08:10:54",
      "content": "<p>Thanks a ton!</p>",
      "rawMarkdown": "Thanks a ton!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2491339,
      "author_name": "hubert101",
      "author_url": "",
      "post_date": "10/21/2023 14:32:39",
      "content": "<p>Interesting approaches! The idea of converting audio to spectrograms is particularly intriguing. Thanks for sharing your experiences – they will surely help others in future projects. Best of luck in upcoming contests!</p>",
      "votes": null,
      "replies": [
        {
          "id": 2492031,
          "author_name": "jainam213",
          "author_url": "",
          "post_date": "10/22/2023 08:10:54",
          "content": "<p>Thanks a ton!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "2489327": "1)I'm surprised none of the high scoring teams converted audio to spectrograms and applied image algorithms on them. Spent quite some time trying this but nothing worked.\n\n2)Next I tried running noise reduction on the audios during infrence. Doesn't add a lot of overhead and noise reduction algorithms work pretty well, this was just making my score worse though. I still think this is a viable option and there should be some(even though slight) improvement especially in OOD domains if we find the sweet spot for noise reduction.\n\n3)What i tried next was implementing/using a linear probe instead of finetuning. So finetuning effects OOD generalization and theoretically linear probes should preform better. (My half baked implementation couldn't though) . I was out of time at this point.\n\nHopefully this helps someone somewhere.",
    "2491339": "Interesting approaches! The idea of converting audio to spectrograms is particularly intriguing. Thanks for sharing your experiences – they will surely help others in future projects. Best of luck in upcoming contests!",
    "2492031": "Thanks a ton!"
  },
  "source": "meta"
}