{
  "id": 426501,
  "title": "Training Time ",
  "url": "/competitions/bengaliai-speech/discussion/426501",
  "author_name": "",
  "post_date": "2023-07-23T19:47:10.346904100Z",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>How long does it take for the participants to train the model<br>\nFor me, it takes 12 hrs per epochs 😪<br>\nHardware - 12GB GPU RTX, 64GB RAM, 20core CPU</p>",
  "messages": [
    {
      "id": "2356021",
      "postDate": "07/23/2023 19:47:10",
      "content": "<p>How long does it take for the participants to train the model<br>\nFor me, it takes 12 hrs per epochs 😪<br>\nHardware - 12GB GPU RTX, 64GB RAM, 20core CPU</p>",
      "rawMarkdown": "How long does it take for the participants to train the model\nFor me, it takes 12 hrs per epochs 😪\nHardware - 12GB GPU RTX, 64GB RAM, 20core CPU",
      "votes": null
    },
    {
      "id": "2360709",
      "postDate": "07/27/2023 00:04:30",
      "content": "<p>it is sufficient to do experiments on subset of the data</p>\n<hr>\n<p>\"In this work, we present an effective scheme of fine-tuning<br>\na pretrained wav2vec2.0 model for speech transcription of a<br>\nlow resource language. We applied the scheme on a subset of<br>\n45 thousand audio samples from the Bengali Common Voice<br>\nDataset to transcribe Bengali audio, achieving a final WER of<br>\n0.2524. This demonstrates the viability of the wav2vec2 model<br>\nfor Bengali Speech Recognition. Furthermore, this result was<br>\nachieved using only 17.84% of the Bengali Common Voices<br>\nDataset making it very likely that even better results can be<br>\nachieved if the entire dataset is utilized. \"</p>\n<p><a href=\"https://arxiv.org/pdf/2209.06581.pdf\" target=\"_blank\">https://arxiv.org/pdf/2209.06581.pdf</a><br>\nApplying wav2vec2 for Speech Recognition on Bengali Common Voices Dataset</p>",
      "rawMarkdown": "it is sufficient to do experiments on subset of the data\n\n----\n\n\"In this work, we present an effective scheme of fine-tuning\na pretrained wav2vec2.0 model for speech transcription of a\nlow resource language. We applied the scheme on a subset of\n45 thousand audio samples from the Bengali Common Voice\nDataset to transcribe Bengali audio, achieving a final WER of\n0.2524. This demonstrates the viability of the wav2vec2 model\nfor Bengali Speech Recognition. Furthermore, this result was\nachieved using only 17.84% of the Bengali Common Voices\nDataset making it very likely that even better results can be\nachieved if the entire dataset is utilized. \"\n\n\nhttps://arxiv.org/pdf/2209.06581.pdf\nApplying wav2vec2 for Speech Recognition on Bengali Common Voices Dataset",
      "votes": null
    },
    {
      "id": "2361814",
      "postDate": "07/27/2023 16:00:03",
      "content": "<p>Thank you 😀 </p>",
      "rawMarkdown": "Thank you 😀",
      "votes": null
    },
    {
      "id": "2379473",
      "postDate": "08/08/2023 07:04:09",
      "content": "<p><a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a> your results in the leadernboard are using subset of data or trained on full data?</p>",
      "rawMarkdown": "hengck23 your results in the leadernboard are using subset of data or trained on full data?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2360709,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "07/27/2023 00:04:30",
      "content": "<p>it is sufficient to do experiments on subset of the data</p>\n<hr>\n<p>\"In this work, we present an effective scheme of fine-tuning<br>\na pretrained wav2vec2.0 model for speech transcription of a<br>\nlow resource language. We applied the scheme on a subset of<br>\n45 thousand audio samples from the Bengali Common Voice<br>\nDataset to transcribe Bengali audio, achieving a final WER of<br>\n0.2524. This demonstrates the viability of the wav2vec2 model<br>\nfor Bengali Speech Recognition. Furthermore, this result was<br>\nachieved using only 17.84% of the Bengali Common Voices<br>\nDataset making it very likely that even better results can be<br>\nachieved if the entire dataset is utilized. \"</p>\n<p><a href=\"https://arxiv.org/pdf/2209.06581.pdf\" target=\"_blank\">https://arxiv.org/pdf/2209.06581.pdf</a><br>\nApplying wav2vec2 for Speech Recognition on Bengali Common Voices Dataset</p>",
      "votes": null,
      "replies": [
        {
          "id": 2361814,
          "author_name": "arunodhayan",
          "author_url": "",
          "post_date": "07/27/2023 16:00:03",
          "content": "<p>Thank you 😀 </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 2379473,
      "author_name": "rajgothi",
      "author_url": "",
      "post_date": "08/08/2023 07:04:09",
      "content": "<p><a href=\"https://www.kaggle.com/hengck23\" target=\"_blank\">@hengck23</a> your results in the leadernboard are using subset of data or trained on full data?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2356021": "How long does it take for the participants to train the model\nFor me, it takes 12 hrs per epochs 😪\nHardware - 12GB GPU RTX, 64GB RAM, 20core CPU",
    "2360709": "it is sufficient to do experiments on subset of the data\n\n----\n\n\"In this work, we present an effective scheme of fine-tuning\na pretrained wav2vec2.0 model for speech transcription of a\nlow resource language. We applied the scheme on a subset of\n45 thousand audio samples from the Bengali Common Voice\nDataset to transcribe Bengali audio, achieving a final WER of\n0.2524. This demonstrates the viability of the wav2vec2 model\nfor Bengali Speech Recognition. Furthermore, this result was\nachieved using only 17.84% of the Bengali Common Voices\nDataset making it very likely that even better results can be\nachieved if the entire dataset is utilized. \"\n\n\nhttps://arxiv.org/pdf/2209.06581.pdf\nApplying wav2vec2 for Speech Recognition on Bengali Common Voices Dataset",
    "2361814": "Thank you 😀",
    "2379473": "hengck23 your results in the leadernboard are using subset of data or trained on full data?"
  },
  "source": "meta"
}