{
  "id": 46497,
  "title": "Silence Class Generation in Validation Set",
  "url": "/competitions/tensorflow-speech-recognition-challenge/discussion/46497",
  "author_name": "",
  "post_date": "2017-12-28T10:17:23.447102300Z",
  "votes": null,
  "comment_count": 3,
  "views": 0,
  "content": "<p>In validation set, I have to generate data for silence class.</p>\n\n<p>My training code includes clipping from _background_noise_ data and augmentations. </p>\n\n<p>Specifically, I'm doing following augmentations.</p>\n\n<ul>\n<li>Noise injection</li>\n<li>Volume Adjustment</li>\n<li>Add Sound effects(using librosa)</li>\n</ul>\n\n<p>But I'm highly suspecious on this strategy, since my model can not predict silence class on test set.\n(I sampled few audios and listened them, and saw my model's predictions.)</p>\n\n<p>Any idea on this?</p>",
  "messages": [
    {
      "id": "262927",
      "postDate": "12/28/2017 10:17:23",
      "content": "<p>In validation set, I have to generate data for silence class.</p>\n\n<p>My training code includes clipping from _background_noise_ data and augmentations. </p>\n\n<p>Specifically, I'm doing following augmentations.</p>\n\n<ul>\n<li>Noise injection</li>\n<li>Volume Adjustment</li>\n<li>Add Sound effects(using librosa)</li>\n</ul>\n\n<p>But I'm highly suspecious on this strategy, since my model can not predict silence class on test set.\n(I sampled few audios and listened them, and saw my model's predictions.)</p>\n\n<p>Any idea on this?</p>",
      "rawMarkdown": "In validation set, I have to generate data for silence class.\n\nMy training code includes clipping from _background_noise_ data and augmentations. \n\nSpecifically, I'm doing following augmentations.\n\n- Noise injection\n- Volume Adjustment\n- Add Sound effects(using librosa)\n\nBut I'm highly suspecious on this strategy, since my model can not predict silence class on test set.\n(I sampled few audios and listened them, and saw my model's predictions.)\n\nAny idea on this?",
      "votes": null
    },
    {
      "id": "262967",
      "postDate": "12/28/2017 13:30:33",
      "content": "<p>If you are talking about misclassified cases on the test set, one thing to remember is that not all test data are evaluated for leaderboard (both private and public). If you adjust your approach to do better on those you observed, while they were actually not evaluated, not sure if you are gaining an edge or actually get worse.</p>",
      "rawMarkdown": "If you are talking about misclassified cases on the test set, one thing to remember is that not all test data are evaluated for leaderboard (both private and public). If you adjust your approach to do better on those you observed, while they were actually not evaluated, not sure if you are gaining an edge or actually get worse.",
      "votes": null
    },
    {
      "id": "263119",
      "postDate": "12/29/2017 00:42:16",
      "content": "<p>Sure. I just want to check what is the weakness in my model and It seems to me that it is very obvious that my model can not predict silence well.. </p>",
      "rawMarkdown": "Sure. I just want to check what is the weakness in my model and It seems to me that it is very obvious that my model can not predict silence well..",
      "votes": null
    },
    {
      "id": "263121",
      "postDate": "12/29/2017 00:54:10",
      "content": "<p>I was able to get better results on silence by supplementing _background_noise with randomly generated white noise, red noise, etc. Also it is possible to use a VAD to detect test set noise and supplement the training data with this, although it is unclear if a VAD is allowed.</p>",
      "rawMarkdown": "I was able to get better results on silence by supplementing _background_noise with randomly generated white noise, red noise, etc. Also it is possible to use a VAD to detect test set noise and supplement the training data with this, although it is unclear if a VAD is allowed.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 262967,
      "author_name": "ryanzhang",
      "author_url": "",
      "post_date": "12/28/2017 13:30:33",
      "content": "<p>If you are talking about misclassified cases on the test set, one thing to remember is that not all test data are evaluated for leaderboard (both private and public). If you adjust your approach to do better on those you observed, while they were actually not evaluated, not sure if you are gaining an edge or actually get worse.</p>",
      "votes": null,
      "replies": [
        {
          "id": 263119,
          "author_name": "ildoonet",
          "author_url": "",
          "post_date": "12/29/2017 00:42:16",
          "content": "<p>Sure. I just want to check what is the weakness in my model and It seems to me that it is very obvious that my model can not predict silence well.. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 263121,
      "author_name": "jfkingiii",
      "author_url": "",
      "post_date": "12/29/2017 00:54:10",
      "content": "<p>I was able to get better results on silence by supplementing _background_noise with randomly generated white noise, red noise, etc. Also it is possible to use a VAD to detect test set noise and supplement the training data with this, although it is unclear if a VAD is allowed.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "262927": "In validation set, I have to generate data for silence class.\n\nMy training code includes clipping from _background_noise_ data and augmentations. \n\nSpecifically, I'm doing following augmentations.\n\n- Noise injection\n- Volume Adjustment\n- Add Sound effects(using librosa)\n\nBut I'm highly suspecious on this strategy, since my model can not predict silence class on test set.\n(I sampled few audios and listened them, and saw my model's predictions.)\n\nAny idea on this?",
    "262967": "If you are talking about misclassified cases on the test set, one thing to remember is that not all test data are evaluated for leaderboard (both private and public). If you adjust your approach to do better on those you observed, while they were actually not evaluated, not sure if you are gaining an edge or actually get worse.",
    "263119": "Sure. I just want to check what is the weakness in my model and It seems to me that it is very obvious that my model can not predict silence well..",
    "263121": "I was able to get better results on silence by supplementing _background_noise with randomly generated white noise, red noise, etc. Also it is possible to use a VAD to detect test set noise and supplement the training data with this, although it is unclear if a VAD is allowed."
  },
  "source": "meta"
}