{
  "id": 44640,
  "title": "Mozilla Open Sourced a Speech Recognition Model Yesterday",
  "url": "/competitions/tensorflow-speech-recognition-challenge/discussion/44640",
  "author_name": "",
  "post_date": "2017-11-30T15:34:21.519704400Z",
  "votes": 4,
  "comment_count": 5,
  "views": 0,
  "content": "<p>The error rate is 6.5%</p>\n\n<p>News: <a href=\"https://blog.mozilla.org/blog/2017/11/29/announcing-the-initial-release-of-mozillas-open-source-speech-recognition-model-and-voice-dataset/\">https://blog.mozilla.org/blog/2017/11/29/announcing-the-initial-release-of-mozillas-open-source-speech-recognition-model-and-voice-dataset/</a></p>\n\n<p>TF code: <a href=\"https://github.com/mozilla/DeepSpeech\">https://github.com/mozilla/DeepSpeech</a></p>",
  "messages": [
    {
      "id": "251008",
      "postDate": "11/30/2017 15:34:21",
      "content": "<p>The error rate is 6.5%</p>\n\n<p>News: <a href=\"https://blog.mozilla.org/blog/2017/11/29/announcing-the-initial-release-of-mozillas-open-source-speech-recognition-model-and-voice-dataset/\">https://blog.mozilla.org/blog/2017/11/29/announcing-the-initial-release-of-mozillas-open-source-speech-recognition-model-and-voice-dataset/</a></p>\n\n<p>TF code: <a href=\"https://github.com/mozilla/DeepSpeech\">https://github.com/mozilla/DeepSpeech</a></p>",
      "rawMarkdown": "The error rate is 6.5%\n\nNews: https://blog.mozilla.org/blog/2017/11/29/announcing-the-initial-release-of-mozillas-open-source-speech-recognition-model-and-voice-dataset/\n\nTF code: https://github.com/mozilla/DeepSpeech",
      "votes": null
    },
    {
      "id": "251146",
      "postDate": "11/30/2017 18:39:52",
      "content": "<p>Thanks for the link.\nToo bad petrained model are not allowed :/</p>",
      "rawMarkdown": "Thanks for the link.\nToo bad petrained model are not allowed :/",
      "votes": null
    },
    {
      "id": "251355",
      "postDate": "12/01/2017 03:36:18",
      "content": "<p>would be interesting to see how much finetuning this would help...  even if we dont submit it </p>",
      "rawMarkdown": "would be interesting to see how much finetuning this would help...  even if we dont submit it",
      "votes": null
    },
    {
      "id": "253395",
      "postDate": "12/04/2017 21:59:57",
      "content": "<p>For fun, I ran the DeepSpeech pretrained model on the train dataset (took ~ 24 hours on a four core desktop CPU...). The recognized texts are here: <a href=\"https://www.kaggle.com/holzner/tf-speechrec-deepspeech-train\">https://www.kaggle.com/holzner/tf-speechrec-deepspeech-train</a> .</p>\n\n<p>And here is a notebook looking at how many of the labels to predict the model got exactly right: <a href=\"https://www.kaggle.com/holzner/deepspeech-predictions\">https://www.kaggle.com/holzner/deepspeech-predictions</a> . The best class is 'yes' where DeepSpeech recognizes 'yes' in for 84% of them.</p>",
      "rawMarkdown": "For fun, I ran the DeepSpeech pretrained model on the train dataset (took ~ 24 hours on a four core desktop CPU...). The recognized texts are here: https://www.kaggle.com/holzner/tf-speechrec-deepspeech-train .\n\nAnd here is a notebook looking at how many of the labels to predict the model got exactly right: https://www.kaggle.com/holzner/deepspeech-predictions . The best class is 'yes' where DeepSpeech recognizes 'yes' in for 84% of them.",
      "votes": null
    },
    {
      "id": "253411",
      "postDate": "12/04/2017 22:34:59",
      "content": "<p>Thanks for sharing!</p>",
      "rawMarkdown": "Thanks for sharing!",
      "votes": null
    },
    {
      "id": "257218",
      "postDate": "12/13/2017 18:14:17",
      "content": "<p>They also made available a voice dataset</p>\n\n<p><a href=\"https://common-voice-data-download.s3.amazonaws.com/cv_corpus_v1.tar.gz\">https://common-voice-data-download.s3.amazonaws.com/cv_corpus_v1.tar.gz</a> (via <a href=\"https://voice.mozilla.org/data\">https://voice.mozilla.org/data</a>)</p>\n\n<p>but I am unable to download it. The download starts but stalls out soon afterwards. Can anyone here verify that this isn't just a problem for me?</p>",
      "rawMarkdown": "They also made available a voice dataset\n\nhttps://common-voice-data-download.s3.amazonaws.com/cv_corpus_v1.tar.gz (via https://voice.mozilla.org/data)\n\nbut I am unable to download it. The download starts but stalls out soon afterwards. Can anyone here verify that this isn't just a problem for me?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 251146,
      "author_name": "CVxTz",
      "author_url": "",
      "post_date": "11/30/2017 18:39:52",
      "content": "<p>Thanks for the link.\nToo bad petrained model are not allowed :/</p>",
      "votes": null,
      "replies": [
        {
          "id": 251355,
          "author_name": "",
          "author_url": "",
          "post_date": "12/01/2017 03:36:18",
          "content": "<p>would be interesting to see how much finetuning this would help...  even if we dont submit it </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 253395,
      "author_name": "holzner",
      "author_url": "",
      "post_date": "12/04/2017 21:59:57",
      "content": "<p>For fun, I ran the DeepSpeech pretrained model on the train dataset (took ~ 24 hours on a four core desktop CPU...). The recognized texts are here: <a href=\"https://www.kaggle.com/holzner/tf-speechrec-deepspeech-train\">https://www.kaggle.com/holzner/tf-speechrec-deepspeech-train</a> .</p>\n\n<p>And here is a notebook looking at how many of the labels to predict the model got exactly right: <a href=\"https://www.kaggle.com/holzner/deepspeech-predictions\">https://www.kaggle.com/holzner/deepspeech-predictions</a> . The best class is 'yes' where DeepSpeech recognizes 'yes' in for 84% of them.</p>",
      "votes": null,
      "replies": [
        {
          "id": 253411,
          "author_name": "shujian",
          "author_url": "",
          "post_date": "12/04/2017 22:34:59",
          "content": "<p>Thanks for sharing!</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 257218,
      "author_name": "innerproduct",
      "author_url": "",
      "post_date": "12/13/2017 18:14:17",
      "content": "<p>They also made available a voice dataset</p>\n\n<p><a href=\"https://common-voice-data-download.s3.amazonaws.com/cv_corpus_v1.tar.gz\">https://common-voice-data-download.s3.amazonaws.com/cv_corpus_v1.tar.gz</a> (via <a href=\"https://voice.mozilla.org/data\">https://voice.mozilla.org/data</a>)</p>\n\n<p>but I am unable to download it. The download starts but stalls out soon afterwards. Can anyone here verify that this isn't just a problem for me?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "251008": "The error rate is 6.5%\n\nNews: https://blog.mozilla.org/blog/2017/11/29/announcing-the-initial-release-of-mozillas-open-source-speech-recognition-model-and-voice-dataset/\n\nTF code: https://github.com/mozilla/DeepSpeech",
    "251146": "Thanks for the link.\nToo bad petrained model are not allowed :/",
    "251355": "would be interesting to see how much finetuning this would help...  even if we dont submit it",
    "253395": "For fun, I ran the DeepSpeech pretrained model on the train dataset (took ~ 24 hours on a four core desktop CPU...). The recognized texts are here: https://www.kaggle.com/holzner/tf-speechrec-deepspeech-train .\n\nAnd here is a notebook looking at how many of the labels to predict the model got exactly right: https://www.kaggle.com/holzner/deepspeech-predictions . The best class is 'yes' where DeepSpeech recognizes 'yes' in for 84% of them.",
    "253411": "Thanks for sharing!",
    "257218": "They also made available a voice dataset\n\nhttps://common-voice-data-download.s3.amazonaws.com/cv_corpus_v1.tar.gz (via https://voice.mozilla.org/data)\n\nbut I am unable to download it. The download starts but stalls out soon afterwards. Can anyone here verify that this isn't just a problem for me?"
  },
  "source": "meta"
}