{
  "id": 442590,
  "title": "WAV and MP3",
  "url": "/competitions/bengaliai-speech/discussion/442590",
  "author_name": "",
  "post_date": "2023-09-23T10:38:59.617744300Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>My model is trained with wav files (are converted from mp3 files). But in prediction phrase with GPU, I see that: </p>\n<ul>\n<li>Results that read from same file by torchaudio and librosa are slightly different. So prediction  result is slight different in infererence phrase. </li>\n<li><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F13344555%2F1e9c1a66b2050561f1932e2f28a95301%2F381401558_1425498424677095_2040793513662364778_n.png?generation=1695465429126342&amp;alt=media\" alt=\"\"></li>\n</ul>\n<p>So are there any ways to make them same when load by librosa without save file(I mean that results are the same when i convert mp3 file to wav file and load wav file with torchaudio)</p>",
  "messages": [
    {
      "id": "2452493",
      "postDate": "09/23/2023 10:38:59",
      "content": "<p>My model is trained with wav files (are converted from mp3 files). But in prediction phrase with GPU, I see that: </p>\n<ul>\n<li>Results that read from same file by torchaudio and librosa are slightly different. So prediction  result is slight different in infererence phrase. </li>\n<li><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F13344555%2F1e9c1a66b2050561f1932e2f28a95301%2F381401558_1425498424677095_2040793513662364778_n.png?generation=1695465429126342&amp;alt=media\" alt=\"\"></li>\n</ul>\n<p>So are there any ways to make them same when load by librosa without save file(I mean that results are the same when i convert mp3 file to wav file and load wav file with torchaudio)</p>",
      "rawMarkdown": "My model is trained with wav files (are converted from mp3 files). But in prediction phrase with GPU, I see that: \n\n+ Results that read from same file by torchaudio and librosa are slightly different. So prediction  result is slight different in infererence phrase. \n+ ![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F13344555%2F1e9c1a66b2050561f1932e2f28a95301%2F381401558_1425498424677095_2040793513662364778_n.png?generation=1695465429126342&alt=media)\n\nSo are there any ways to make them same when load by librosa without save file(I mean that results are the same when i convert mp3 file to wav file and load wav file with torchaudio)",
      "votes": null
    },
    {
      "id": "2452760",
      "postDate": "09/23/2023 15:00:15",
      "content": "<p>Can you plot the difference between wave1 and wave2 ? If they are extremely small, it's probably just the precision of mp3 decoder.<br>\nMp3 is a lossy compression</p>",
      "rawMarkdown": "Can you plot the difference between wave1 and wave2 ? If they are extremely small, it's probably just the precision of mp3 decoder.\nMp3 is a lossy compression",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2452760,
      "author_name": "nyleve",
      "author_url": "",
      "post_date": "09/23/2023 15:00:15",
      "content": "<p>Can you plot the difference between wave1 and wave2 ? If they are extremely small, it's probably just the precision of mp3 decoder.<br>\nMp3 is a lossy compression</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2452493": "My model is trained with wav files (are converted from mp3 files). But in prediction phrase with GPU, I see that: \n\n+ Results that read from same file by torchaudio and librosa are slightly different. So prediction  result is slight different in infererence phrase. \n+ ![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F13344555%2F1e9c1a66b2050561f1932e2f28a95301%2F381401558_1425498424677095_2040793513662364778_n.png?generation=1695465429126342&alt=media)\n\nSo are there any ways to make them same when load by librosa without save file(I mean that results are the same when i convert mp3 file to wav file and load wav file with torchaudio)",
    "2452760": "Can you plot the difference between wave1 and wave2 ? If they are extremely small, it's probably just the precision of mp3 decoder.\nMp3 is a lossy compression"
  },
  "source": "meta"
}