{
  "id": 176968,
  "title": "Timestamps for train data",
  "url": "/competitions/birdsong-recognition/discussion/176968",
  "author_name": "",
  "post_date": "2020-08-24T09:24:43.972905500Z",
  "votes": 5,
  "comment_count": 4,
  "views": 0,
  "content": "<p>I found some time and marked up most of the sounds. (posted in the <a href=\"https://www.kaggle.com/sapr3s/timestamps-for-train-data-cornell-birdcall\" target=\"_blank\">dataset</a>) </p>\n<p>The markup consists of the following data (time in seconds):</p>\n<table>\n<thead>\n<tr>\n<th>bird</th>\n<th>file</th>\n<th>start</th>\n<th>duration</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>aldfly</td>\n<td>XC134874.mp3</td>\n<td>4.7</td>\n<td>0.4</td>\n</tr>\n<tr>\n<td>aldfly</td>\n<td>XC134874.mp3</td>\n<td>8.4</td>\n<td>0.4</td>\n</tr>\n<tr>\n<td>…</td>\n<td>…</td>\n<td>…</td>\n<td>…</td>\n</tr>\n</tbody>\n</table>\n<p>Schematically, the markup principle is shown at the image <img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1275752%2Fc852476b6723567209c9767238353727%2Ftimestamps.png?generation=1598260437041116&amp;alt=media\" alt=\"figure\"></p>\n<p><strong>Known features</strong></p>\n<ul>\n<li>time was rounded up to 0.1 seconds</li>\n<li>(1 in the fig) some weak signals were not included in the markup</li>\n<li>(2 in the fig) signals from other classes or \"nocall\" are included in the markup (voices, etc.)</li>\n<li>(3 in the fig) some sequences of sounds (\"words in a song\") are marked as separate sounds</li>\n<li>(4 in the fig) other sequences (\"words in a song\") are marked in one line</li>\n<li>in the markup got 20225 files from 21375 (1150 files skipped)</li>\n</ul>\n<p>Of course, I haven't checked everything yet. Therefore, if there are serious errors (other than those indicated above), then please report.</p>\n<p>Small <a href=\"https://www.kaggle.com/sapr3s/timestamps-example\" target=\"_blank\">example</a> of use</p>\n<p><strong>TODO</strong>  (as far as possible)</p>\n<ul>\n<li>remove unnecessary sounds that do not belong to this class (using a trained classifier)</li>\n<li>add weak sounds of this class that are not included in the markup</li>\n<li>deal with some classes that are not well marked up (for example, I didn't like the \"amebit\" markup)</li>\n</ul>\n<hr>\n<p>As you know: the more accurate the marking - the more accurate the model.<br>\nIt's time to take the 0.7 on LB! <br>\nWhat do you think?</p>",
  "messages": [
    {
      "id": "983402",
      "postDate": "08/24/2020 09:24:43",
      "content": "<p>I found some time and marked up most of the sounds. (posted in the <a href=\"https://www.kaggle.com/sapr3s/timestamps-for-train-data-cornell-birdcall\" target=\"_blank\">dataset</a>) </p>\n<p>The markup consists of the following data (time in seconds):</p>\n<table>\n<thead>\n<tr>\n<th>bird</th>\n<th>file</th>\n<th>start</th>\n<th>duration</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>aldfly</td>\n<td>XC134874.mp3</td>\n<td>4.7</td>\n<td>0.4</td>\n</tr>\n<tr>\n<td>aldfly</td>\n<td>XC134874.mp3</td>\n<td>8.4</td>\n<td>0.4</td>\n</tr>\n<tr>\n<td>…</td>\n<td>…</td>\n<td>…</td>\n<td>…</td>\n</tr>\n</tbody>\n</table>\n<p>Schematically, the markup principle is shown at the image <img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1275752%2Fc852476b6723567209c9767238353727%2Ftimestamps.png?generation=1598260437041116&amp;alt=media\" alt=\"figure\"></p>\n<p><strong>Known features</strong></p>\n<ul>\n<li>time was rounded up to 0.1 seconds</li>\n<li>(1 in the fig) some weak signals were not included in the markup</li>\n<li>(2 in the fig) signals from other classes or \"nocall\" are included in the markup (voices, etc.)</li>\n<li>(3 in the fig) some sequences of sounds (\"words in a song\") are marked as separate sounds</li>\n<li>(4 in the fig) other sequences (\"words in a song\") are marked in one line</li>\n<li>in the markup got 20225 files from 21375 (1150 files skipped)</li>\n</ul>\n<p>Of course, I haven't checked everything yet. Therefore, if there are serious errors (other than those indicated above), then please report.</p>\n<p>Small <a href=\"https://www.kaggle.com/sapr3s/timestamps-example\" target=\"_blank\">example</a> of use</p>\n<p><strong>TODO</strong>  (as far as possible)</p>\n<ul>\n<li>remove unnecessary sounds that do not belong to this class (using a trained classifier)</li>\n<li>add weak sounds of this class that are not included in the markup</li>\n<li>deal with some classes that are not well marked up (for example, I didn't like the \"amebit\" markup)</li>\n</ul>\n<hr>\n<p>As you know: the more accurate the marking - the more accurate the model.<br>\nIt's time to take the 0.7 on LB! <br>\nWhat do you think?</p>",
      "rawMarkdown": "I found some time and marked up most of the sounds. (posted in the [dataset](https://www.kaggle.com/sapr3s/timestamps-for-train-data-cornell-birdcall)) \n\nThe markup consists of the following data (time in seconds):\n| bird  | file  |start  | duration |\n| --- | --- |--- |--- |\n|aldfly| XC134874.mp3 | 4.7 | 0.4 |\n|aldfly| XC134874.mp3 | 8.4 | 0.4 |\n| ... | ... | ... | ... |\n\n\nSchematically, the markup principle is shown at the image ![figure](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1275752%2Fc852476b6723567209c9767238353727%2Ftimestamps.png?generation=1598260437041116&alt=media)\n\n**Known features**\n- time was rounded up to 0.1 seconds\n- (1 in the fig) some weak signals were not included in the markup\n- (2 in the fig) signals from other classes or \"nocall\" are included in the markup (voices, etc.)\n- (3 in the fig) some sequences of sounds (\"words in a song\") are marked as separate sounds\n- (4 in the fig) other sequences (\"words in a song\") are marked in one line\n- in the markup got 20225 files from 21375 (1150 files skipped)\n\nOf course, I haven't checked everything yet. Therefore, if there are serious errors (other than those indicated above), then please report.\n\nSmall [example](https://www.kaggle.com/sapr3s/timestamps-example) of use\n\n**TODO**  (as far as possible)\n- remove unnecessary sounds that do not belong to this class (using a trained classifier)\n- add weak sounds of this class that are not included in the markup\n- deal with some classes that are not well marked up (for example, I didn't like the \"amebit\" markup)\n\n\n---\nAs you know: the more accurate the marking - the more accurate the model.\nIt's time to take the 0.7 on LB! \nWhat do you think?",
      "votes": null
    },
    {
      "id": "983515",
      "postDate": "08/24/2020 11:21:30",
      "content": "<p>I do not completely understand. Is this markup to get only \"singed\" pieces of audio?</p>",
      "rawMarkdown": "I do not completely understand. Is this markup to get only \"singed\" pieces of audio?",
      "votes": null
    },
    {
      "id": "983557",
      "postDate": "08/24/2020 12:03:23",
      "content": "<p>I tried to find sounds that are somehow different from background noise. And filtered by level. There could also be other loud sounds that are not 'noise' and are not 'birdsong' (voices, etc.) … Ideally, you need an additional classification</p>",
      "rawMarkdown": "I tried to find sounds that are somehow different from background noise. And filtered by level. There could also be other loud sounds that are not 'noise' and are not 'birdsong' (voices, etc.) ... Ideally, you need an additional classification",
      "votes": null
    },
    {
      "id": "1001559",
      "postDate": "09/07/2020 12:05:54",
      "content": "<p>How did you create the mark?  Thresholding on sound level?</p>",
      "rawMarkdown": "How did you create the mark?  Thresholding on sound level?",
      "votes": null
    },
    {
      "id": "1001759",
      "postDate": "09/07/2020 14:39:01",
      "content": "<p>Threshold value, but not to pure sound, but after some conversion. We can say that these are frequency levels and their deviations. But there were a lot of unnecessary sounds. When I checked the first examples, about 10% should be discarded…. It's funny, but I was able to separate sounds better than classify them) Noise greatly distorts the classification result</p>",
      "rawMarkdown": "Threshold value, but not to pure sound, but after some conversion. We can say that these are frequency levels and their deviations. But there were a lot of unnecessary sounds. When I checked the first examples, about 10% should be discarded.... It's funny, but I was able to separate sounds better than classify them) Noise greatly distorts the classification result",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 983515,
      "author_name": "vladimirsydor",
      "author_url": "",
      "post_date": "08/24/2020 11:21:30",
      "content": "<p>I do not completely understand. Is this markup to get only \"singed\" pieces of audio?</p>",
      "votes": null,
      "replies": [
        {
          "id": 983557,
          "author_name": "sapr3s",
          "author_url": "",
          "post_date": "08/24/2020 12:03:23",
          "content": "<p>I tried to find sounds that are somehow different from background noise. And filtered by level. There could also be other loud sounds that are not 'noise' and are not 'birdsong' (voices, etc.) … Ideally, you need an additional classification</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 1001559,
      "author_name": "cpmpml",
      "author_url": "",
      "post_date": "09/07/2020 12:05:54",
      "content": "<p>How did you create the mark?  Thresholding on sound level?</p>",
      "votes": null,
      "replies": [
        {
          "id": 1001759,
          "author_name": "sapr3s",
          "author_url": "",
          "post_date": "09/07/2020 14:39:01",
          "content": "<p>Threshold value, but not to pure sound, but after some conversion. We can say that these are frequency levels and their deviations. But there were a lot of unnecessary sounds. When I checked the first examples, about 10% should be discarded…. It's funny, but I was able to separate sounds better than classify them) Noise greatly distorts the classification result</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "983402": "I found some time and marked up most of the sounds. (posted in the [dataset](https://www.kaggle.com/sapr3s/timestamps-for-train-data-cornell-birdcall)) \n\nThe markup consists of the following data (time in seconds):\n| bird  | file  |start  | duration |\n| --- | --- |--- |--- |\n|aldfly| XC134874.mp3 | 4.7 | 0.4 |\n|aldfly| XC134874.mp3 | 8.4 | 0.4 |\n| ... | ... | ... | ... |\n\n\nSchematically, the markup principle is shown at the image ![figure](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1275752%2Fc852476b6723567209c9767238353727%2Ftimestamps.png?generation=1598260437041116&alt=media)\n\n**Known features**\n- time was rounded up to 0.1 seconds\n- (1 in the fig) some weak signals were not included in the markup\n- (2 in the fig) signals from other classes or \"nocall\" are included in the markup (voices, etc.)\n- (3 in the fig) some sequences of sounds (\"words in a song\") are marked as separate sounds\n- (4 in the fig) other sequences (\"words in a song\") are marked in one line\n- in the markup got 20225 files from 21375 (1150 files skipped)\n\nOf course, I haven't checked everything yet. Therefore, if there are serious errors (other than those indicated above), then please report.\n\nSmall [example](https://www.kaggle.com/sapr3s/timestamps-example) of use\n\n**TODO**  (as far as possible)\n- remove unnecessary sounds that do not belong to this class (using a trained classifier)\n- add weak sounds of this class that are not included in the markup\n- deal with some classes that are not well marked up (for example, I didn't like the \"amebit\" markup)\n\n\n---\nAs you know: the more accurate the marking - the more accurate the model.\nIt's time to take the 0.7 on LB! \nWhat do you think?",
    "983515": "I do not completely understand. Is this markup to get only \"singed\" pieces of audio?",
    "983557": "I tried to find sounds that are somehow different from background noise. And filtered by level. There could also be other loud sounds that are not 'noise' and are not 'birdsong' (voices, etc.) ... Ideally, you need an additional classification",
    "1001559": "How did you create the mark?  Thresholding on sound level?",
    "1001759": "Threshold value, but not to pure sound, but after some conversion. We can say that these are frequency levels and their deviations. But there were a lot of unnecessary sounds. When I checked the first examples, about 10% should be discarded.... It's funny, but I was able to separate sounds better than classify them) Noise greatly distorts the classification result"
  },
  "source": "meta"
}