{
  "id": 45547,
  "title": "High level Google keyword spotting architecture",
  "url": "/competitions/tensorflow-speech-recognition-challenge/discussion/45547",
  "author_name": "",
  "post_date": "2017-12-12T18:35:18.678210200Z",
  "votes": 7,
  "comment_count": 3,
  "views": 0,
  "content": "<p>A paper we just presented at NIPS about our keyword (aka \"OK Google\") spotting system.\n<a href=\"http://arxiv.org/abs/1712.03603\">http://arxiv.org/abs/1712.03603</a></p>\n\n<p>The paper is very high level, but links to relevant past papers.</p>",
  "messages": [
    {
      "id": "256796",
      "postDate": "12/12/2017 18:35:18",
      "content": "<p>A paper we just presented at NIPS about our keyword (aka \"OK Google\") spotting system.\n<a href=\"http://arxiv.org/abs/1712.03603\">http://arxiv.org/abs/1712.03603</a></p>\n\n<p>The paper is very high level, but links to relevant past papers.</p>",
      "rawMarkdown": "A paper we just presented at NIPS about our keyword (aka \"OK Google\") spotting system.\nhttp://arxiv.org/abs/1712.03603\n\nThe paper is very high level, but links to relevant past papers.",
      "votes": null
    },
    {
      "id": "257062",
      "postDate": "12/13/2017 09:59:52",
      "content": "<p>Here's a similar article for \"Hey Siri\" from Apple: <a href=\"https://machinelearning.apple.com/2017/10/01/hey-siri.html\">https://machinelearning.apple.com/2017/10/01/hey-siri.html</a></p>",
      "rawMarkdown": "Here's a similar article for \"Hey Siri\" from Apple: https://machinelearning.apple.com/2017/10/01/hey-siri.html",
      "votes": null
    },
    {
      "id": "261687",
      "postDate": "12/23/2017 15:16:21",
      "content": "<p>I thought your system worked more like <a href=\"https://www.clsp.jhu.edu/wp-content/uploads/sites/75/2015/11/icassp2015_myhotword.pdf\">https://www.clsp.jhu.edu/wp-content/uploads/sites/75/2015/11/icassp2015_myhotword.pdf</a> Crucially, you first train a word-level ASR model that really \"gets\" speech, and then use it to spot keywords. This dataset is quite information-poor (very little variety in negative examples) compared to a usual speech corpus.</p>",
      "rawMarkdown": "I thought your system worked more like https://www.clsp.jhu.edu/wp-content/uploads/sites/75/2015/11/icassp2015_myhotword.pdf Crucially, you first train a word-level ASR model that really \"gets\" speech, and then use it to spot keywords. This dataset is quite information-poor (very little variety in negative examples) compared to a usual speech corpus.",
      "votes": null
    },
    {
      "id": "261776",
      "postDate": "12/23/2017 23:21:29",
      "content": "<p>That was an attempt at building a detector for user defined keywords. This contest as well as all commercial assistants have a statically defined keyword(s).</p>",
      "rawMarkdown": "That was an attempt at building a detector for user defined keywords. This contest as well as all commercial assistants have a statically defined keyword(s).",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 257062,
      "author_name": "humananalog",
      "author_url": "",
      "post_date": "12/13/2017 09:59:52",
      "content": "<p>Here's a similar article for \"Hey Siri\" from Apple: <a href=\"https://machinelearning.apple.com/2017/10/01/hey-siri.html\">https://machinelearning.apple.com/2017/10/01/hey-siri.html</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 261687,
      "author_name": "adubinsky",
      "author_url": "",
      "post_date": "12/23/2017 15:16:21",
      "content": "<p>I thought your system worked more like <a href=\"https://www.clsp.jhu.edu/wp-content/uploads/sites/75/2015/11/icassp2015_myhotword.pdf\">https://www.clsp.jhu.edu/wp-content/uploads/sites/75/2015/11/icassp2015_myhotword.pdf</a> Crucially, you first train a word-level ASR model that really \"gets\" speech, and then use it to spot keywords. This dataset is quite information-poor (very little variety in negative examples) compared to a usual speech corpus.</p>",
      "votes": null,
      "replies": [
        {
          "id": 261776,
          "author_name": "razielalvarez",
          "author_url": "",
          "post_date": "12/23/2017 23:21:29",
          "content": "<p>That was an attempt at building a detector for user defined keywords. This contest as well as all commercial assistants have a statically defined keyword(s).</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "256796": "A paper we just presented at NIPS about our keyword (aka \"OK Google\") spotting system.\nhttp://arxiv.org/abs/1712.03603\n\nThe paper is very high level, but links to relevant past papers.",
    "257062": "Here's a similar article for \"Hey Siri\" from Apple: https://machinelearning.apple.com/2017/10/01/hey-siri.html",
    "261687": "I thought your system worked more like https://www.clsp.jhu.edu/wp-content/uploads/sites/75/2015/11/icassp2015_myhotword.pdf Crucially, you first train a word-level ASR model that really \"gets\" speech, and then use it to spot keywords. This dataset is quite information-poor (very little variety in negative examples) compared to a usual speech corpus.",
    "261776": "That was an attempt at building a detector for user defined keywords. This contest as well as all commercial assistants have a statically defined keyword(s)."
  },
  "source": "meta"
}