{
  "id": 242714,
  "title": "Has anyone tried nnAudio? How`s its performance vs torchlibrosa?",
  "url": "/competitions/birdclef-2021/discussion/242714",
  "author_name": "Hao",
  "post_date": "2021-05-30T11:58:55.124000",
  "votes": 6,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Just found there is another option to do STFT and LogMel on the fly: <a href=\"https://github.com/KinWaiCheuk/nnAudio\" target=\"_blank\">nnAudio</a>. </p>\n<p>We know usually many people pre-compute STFT and LogMel and save on disk to save training time, but with these saved spectrogram we lost many option/flexibility to do audio augmentation. </p>\n<p>And we tried torchlibrosa but it is still slow, do not know whether nnAudio is better in speed and performance?</p>",
  "messages": [
    {
      "id": 1328601,
      "postDate": "2021-05-30T11:58:55.123Z",
      "content": "<p>Just found there is another option to do STFT and LogMel on the fly: <a href=\"https://github.com/KinWaiCheuk/nnAudio\" target=\"_blank\">nnAudio</a>. </p>\n<p>We know usually many people pre-compute STFT and LogMel and save on disk to save training time, but with these saved spectrogram we lost many option/flexibility to do audio augmentation. </p>\n<p>And we tried torchlibrosa but it is still slow, do not know whether nnAudio is better in speed and performance?</p>",
      "rawMarkdown": "Just found there is another option to do STFT and LogMel on the fly: [nnAudio](https://github.com/KinWaiCheuk/nnAudio). \n\nWe know usually many people pre-compute STFT and LogMel and save on disk to save training time, but with these saved spectrogram we lost many option/flexibility to do audio augmentation. \n\nAnd we tried torchlibrosa but it is still slow, do not know whether nnAudio is better in speed and performance?",
      "votes": 6
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1328601": "Just found there is another option to do STFT and LogMel on the fly: [nnAudio](https://github.com/KinWaiCheuk/nnAudio). \n\nWe know usually many people pre-compute STFT and LogMel and save on disk to save training time, but with these saved spectrogram we lost many option/flexibility to do audio augmentation. \n\nAnd we tried torchlibrosa but it is still slow, do not know whether nnAudio is better in speed and performance?"
  }
}