{
  "id": 133721,
  "title": "Has anyone got benefit from audio?",
  "url": "/competitions/deepfake-detection-challenge/discussion/133721",
  "author_name": "",
  "post_date": "2020-03-04T01:46:38.876216100Z",
  "votes": 4,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I've tried lots of methods (E.g. spectrogram + CNN), but nothing improved. I wonder is audio important here. Has anyone got benefit from audio?</p>",
  "messages": [
    {
      "id": "762974",
      "postDate": "03/04/2020 01:46:38",
      "content": "<p>I've tried lots of methods (E.g. spectrogram + CNN), but nothing improved. I wonder is audio important here. Has anyone got benefit from audio?</p>",
      "rawMarkdown": "I've tried lots of methods (E.g. spectrogram + CNN), but nothing improved. I wonder is audio important here. Has anyone got benefit from audio?",
      "votes": null
    },
    {
      "id": "763209",
      "postDate": "03/04/2020 08:49:10",
      "content": "<p>I am planning to try\n<a href=\"https://github.com/tyiannak/pyAudioAnalysis\">https://github.com/tyiannak/pyAudioAnalysis</a>\n<a href=\"https://github.com/amsehili/auditok\">https://github.com/amsehili/auditok</a></p>",
      "rawMarkdown": "I am planning to try\nhttps://github.com/tyiannak/pyAudioAnalysis\nhttps://github.com/amsehili/auditok",
      "votes": null
    },
    {
      "id": "763964",
      "postDate": "03/05/2020 03:11:14",
      "content": "<p>I was thinking that some of the \"altered\" audio could just be an audio track from another video transposed onto a video with an altered face. That would mean that the audio track itself is unaltered (i.e. you won't see anything on spectrograms etc.) but things like lip sync would be off, which would require some pretty sophisticated models to spot.</p>\n\n<p>One thing that has already been mentioned is the 16Khz audio tracks belonging to fake videos in train, but that \"feature\" may not exist in private test</p>",
      "rawMarkdown": "I was thinking that some of the \"altered\" audio could just be an audio track from another video transposed onto a video with an altered face. That would mean that the audio track itself is unaltered (i.e. you won't see anything on spectrograms etc.) but things like lip sync would be off, which would require some pretty sophisticated models to spot.\n\nOne thing that has already been mentioned is the 16Khz audio tracks belonging to fake videos in train, but that \"feature\" may not exist in private test",
      "votes": null
    },
    {
      "id": "763996",
      "postDate": "03/05/2020 04:05:22",
      "content": "<p>&gt; One thing that has already been mentioned is the 16Khz audio tracks belonging to fake videos in train, but that \"feature\" may not exist in private test</p>\n\n<p>Maybe you're right. I don't know the Private Test Set, but there is no video with 16Khz audio tracks in Public Test Set. </p>",
      "rawMarkdown": "&gt; One thing that has already been mentioned is the 16Khz audio tracks belonging to fake videos in train, but that \"feature\" may not exist in private test\n\nMaybe you're right. I don't know the Private Test Set, but there is no video with 16Khz audio tracks in Public Test Set.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 763209,
      "author_name": "sirishks",
      "author_url": "",
      "post_date": "03/04/2020 08:49:10",
      "content": "<p>I am planning to try\n<a href=\"https://github.com/tyiannak/pyAudioAnalysis\">https://github.com/tyiannak/pyAudioAnalysis</a>\n<a href=\"https://github.com/amsehili/auditok\">https://github.com/amsehili/auditok</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 763964,
      "author_name": "anjum48",
      "author_url": "",
      "post_date": "03/05/2020 03:11:14",
      "content": "<p>I was thinking that some of the \"altered\" audio could just be an audio track from another video transposed onto a video with an altered face. That would mean that the audio track itself is unaltered (i.e. you won't see anything on spectrograms etc.) but things like lip sync would be off, which would require some pretty sophisticated models to spot.</p>\n\n<p>One thing that has already been mentioned is the 16Khz audio tracks belonging to fake videos in train, but that \"feature\" may not exist in private test</p>",
      "votes": null,
      "replies": [
        {
          "id": 763996,
          "author_name": "wuliaokaola",
          "author_url": "",
          "post_date": "03/05/2020 04:05:22",
          "content": "<p>&gt; One thing that has already been mentioned is the 16Khz audio tracks belonging to fake videos in train, but that \"feature\" may not exist in private test</p>\n\n<p>Maybe you're right. I don't know the Private Test Set, but there is no video with 16Khz audio tracks in Public Test Set. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "762974": "I've tried lots of methods (E.g. spectrogram + CNN), but nothing improved. I wonder is audio important here. Has anyone got benefit from audio?",
    "763209": "I am planning to try\nhttps://github.com/tyiannak/pyAudioAnalysis\nhttps://github.com/amsehili/auditok",
    "763964": "I was thinking that some of the \"altered\" audio could just be an audio track from another video transposed onto a video with an altered face. That would mean that the audio track itself is unaltered (i.e. you won't see anything on spectrograms etc.) but things like lip sync would be off, which would require some pretty sophisticated models to spot.\n\nOne thing that has already been mentioned is the 16Khz audio tracks belonging to fake videos in train, but that \"feature\" may not exist in private test",
    "763996": "&gt; One thing that has already been mentioned is the 16Khz audio tracks belonging to fake videos in train, but that \"feature\" may not exist in private test\n\nMaybe you're right. I don't know the Private Test Set, but there is no video with 16Khz audio tracks in Public Test Set."
  },
  "source": "meta"
}