{
  "id": 44218,
  "title": "Silence vs Unknown",
  "url": "/competitions/tensorflow-speech-recognition-challenge/discussion/44218",
  "author_name": "",
  "post_date": "2017-11-25T11:01:07.830464100Z",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Browsing the testset it looks like there are a lot off samples that qualify for both unknown or silence. Some of the samples are really low level, mostly noise.  What defines silence ? My model really struggles on that with practically no \"silence\" samples in the train set to train on.</p>",
  "messages": [
    {
      "id": "248244",
      "postDate": "11/25/2017 11:01:07",
      "content": "<p>Browsing the testset it looks like there are a lot off samples that qualify for both unknown or silence. Some of the samples are really low level, mostly noise.  What defines silence ? My model really struggles on that with practically no \"silence\" samples in the train set to train on.</p>",
      "rawMarkdown": "Browsing the testset it looks like there are a lot off samples that qualify for both unknown or silence. Some of the samples are really low level, mostly noise.  What defines silence ? My model really struggles on that with practically no \"silence\" samples in the train set to train on.",
      "votes": null
    },
    {
      "id": "248287",
      "postDate": "11/25/2017 14:56:06",
      "content": "<p>They have explained in this page ('Unknown Class' and 'Silence'): <a href=\"https://www.tensorflow.org/versions/master/tutorials/audio_recognition\">https://www.tensorflow.org/versions/master/tutorials/audio_recognition</a></p>",
      "rawMarkdown": "They have explained in this page ('Unknown Class' and 'Silence'): https://www.tensorflow.org/versions/master/tutorials/audio_recognition",
      "votes": null
    },
    {
      "id": "248529",
      "postDate": "11/26/2017 11:30:14",
      "content": "<p>What could be an alternative is using a voice detection algorithm first (also useful to cut the silence before the voice starts). </p>\n\n<p>If the voice detection algorithm doesn't find any speech, then of course you could already labeled it silence and wouldn't have to pass to your model at all. So you have to learn only \"unknowns\" in your model classes and no silences anymore.</p>\n\n<p>The downside is that a voice detection algorithm would't meet the rules for the Raspberry Pi competition where you are allowed to only use TensorFlow. </p>",
      "rawMarkdown": "What could be an alternative is using a voice detection algorithm first (also useful to cut the silence before the voice starts). \n\nIf the voice detection algorithm doesn't find any speech, then of course you could already labeled it silence and wouldn't have to pass to your model at all. So you have to learn only \"unknowns\" in your model classes and no silences anymore.\n\nThe downside is that a voice detection algorithm would't meet the rules for the Raspberry Pi competition where you are allowed to only use TensorFlow.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 248287,
      "author_name": "shujian",
      "author_url": "",
      "post_date": "11/25/2017 14:56:06",
      "content": "<p>They have explained in this page ('Unknown Class' and 'Silence'): <a href=\"https://www.tensorflow.org/versions/master/tutorials/audio_recognition\">https://www.tensorflow.org/versions/master/tutorials/audio_recognition</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 248529,
      "author_name": "peterdekkers101",
      "author_url": "",
      "post_date": "11/26/2017 11:30:14",
      "content": "<p>What could be an alternative is using a voice detection algorithm first (also useful to cut the silence before the voice starts). </p>\n\n<p>If the voice detection algorithm doesn't find any speech, then of course you could already labeled it silence and wouldn't have to pass to your model at all. So you have to learn only \"unknowns\" in your model classes and no silences anymore.</p>\n\n<p>The downside is that a voice detection algorithm would't meet the rules for the Raspberry Pi competition where you are allowed to only use TensorFlow. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "248244": "Browsing the testset it looks like there are a lot off samples that qualify for both unknown or silence. Some of the samples are really low level, mostly noise.  What defines silence ? My model really struggles on that with practically no \"silence\" samples in the train set to train on.",
    "248287": "They have explained in this page ('Unknown Class' and 'Silence'): https://www.tensorflow.org/versions/master/tutorials/audio_recognition",
    "248529": "What could be an alternative is using a voice detection algorithm first (also useful to cut the silence before the voice starts). \n\nIf the voice detection algorithm doesn't find any speech, then of course you could already labeled it silence and wouldn't have to pass to your model at all. So you have to learn only \"unknowns\" in your model classes and no silences anymore.\n\nThe downside is that a voice detection algorithm would't meet the rules for the Raspberry Pi competition where you are allowed to only use TensorFlow."
  },
  "source": "meta"
}