{ "id": 211585, "title": "How do I crop a spectrogram/MFCC", "url": "/competitions/rfcx-species-audio-detection/discussion/211585", "author_name": "", "post_date": "2021-01-15T17:59:02.682773300Z", "votes": 1, "comment_count": 3, "views": 0, "content": "

After studying the data a bit I realised that a single audio file has multiple bird sounds as pointed out by @kneroma .
\nBut I am still confused as to how does one crop audio? Should I create an image of all the spectrograms and then use the image co-ordinates to make the crops? Can I do it directly on the spectrogram without converting the spectrogram to an image? If yes then how?

\n

Heres the bounding boxes I want to crop.
\n\"\"

", "messages": [ { "id": "1154573", "postDate": "01/15/2021 17:59:02", "content": "

After studying the data a bit I realised that a single audio file has multiple bird sounds as pointed out by @kneroma .
\nBut I am still confused as to how does one crop audio? Should I create an image of all the spectrograms and then use the image co-ordinates to make the crops? Can I do it directly on the spectrogram without converting the spectrogram to an image? If yes then how?

\n

Heres the bounding boxes I want to crop.
\n\"\"

", "rawMarkdown": "After studying the data a bit I realised that a single audio file has multiple bird sounds as pointed out by @kneroma .\nBut I am still confused as to how does one crop audio? Should I create an image of all the spectrograms and then use the image co-ordinates to make the crops? Can I do it directly on the spectrogram without converting the spectrogram to an image? If yes then how?\n\nHeres the bounding boxes I want to crop.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1604765%2F04415575fe06bdcc94041eed48acdf1f%2FScreenshot%20from%202021-01-15%2023-26-07.png?generation=1610733598342597&alt=media)", "votes": null }, { "id": "1154651", "postDate": "01/15/2021 18:44:18", "content": "

The spectogram is a 2D numpy array that is easily subscriptable. For example, my code to isolate a section with a midpoint (in seconds) and a fixed width:

\n
def sec2index(s, spec_len=5626):\n    return int(np.round(s*spec_len/60))\n\ndef make_slice(mid, width, spec_len):\n    end = mid + width\n    start = mid - width\n\n    if start < 0:\n        end = end - start\n        start = 0\n\n    if end > 60:\n        start = start - (end-60)\n        end = 60\n\n    start_idx = sec2index(start, spec_len)\n    end_idx = sec2index(end, spec_len)\n\n    return start_idx, end_idx\n\nmel_spec = librosa.feature.melspectrogram(audio_vec, n_fft=FFT, hop_length=HOP, sr=SR, fmin=F_MIN, fmax=F_MAX, power=POWER)\n\nspec_len = mel_spec.shape[1]\n\nstart, end = make_slice(t_mid, WINDOW_WIDTH, spec_len)\n\nmel_spec_slice = mel_spec[:, start:end]\n
\n

I'm sure you can modify this code to work on the y-axis/frequency axis as well.

", "rawMarkdown": "The spectogram is a 2D numpy array that is easily subscriptable. For example, my code to isolate a section with a midpoint (in seconds) and a fixed width:\n\n```\ndef sec2index(s, spec_len=5626):\n return int(np.round(s*spec_len/60))\n\ndef make_slice(mid, width, spec_len):\n end = mid + width\n start = mid - width\n \n if start < 0:\n end = end - start\n start = 0\n \n if end > 60:\n start = start - (end-60)\n end = 60\n \n start_idx = sec2index(start, spec_len)\n end_idx = sec2index(end, spec_len)\n \n return start_idx, end_idx\n\nmel_spec = librosa.feature.melspectrogram(audio_vec, n_fft=FFT, hop_length=HOP, sr=SR, fmin=F_MIN, fmax=F_MAX, power=POWER)\n\nspec_len = mel_spec.shape[1]\n\nstart, end = make_slice(t_mid, WINDOW_WIDTH, spec_len)\n\nmel_spec_slice = mel_spec[:, start:end]\n```\n\nI'm sure you can modify this code to work on the y-axis/frequency axis as well.", "votes": null }, { "id": "1156822", "postDate": "01/17/2021 12:37:43", "content": "

Thanks for comment @RNA, but for some reason the same logic doesnt really translate to the y axis :/

", "rawMarkdown": "Thanks for comment @RNA, but for some reason the same logic doesnt really translate to the y axis :/", "votes": null }, { "id": "1156939", "postDate": "01/17/2021 14:38:01", "content": "

I solved it by limiting the librosa mel filter bank to the adequate freq limits, in case any one runs into the same problem

", "rawMarkdown": "I solved it by limiting the librosa mel filter bank to the adequate freq limits, in case any one runs into the same problem", "votes": null } ], "comments": [ { "id": 1154651, "author_name": "bigironsphere", "author_url": "", "post_date": "01/15/2021 18:44:18", "content": "

The spectogram is a 2D numpy array that is easily subscriptable. For example, my code to isolate a section with a midpoint (in seconds) and a fixed width:

\n
def sec2index(s, spec_len=5626):\n    return int(np.round(s*spec_len/60))\n\ndef make_slice(mid, width, spec_len):\n    end = mid + width\n    start = mid - width\n\n    if start < 0:\n        end = end - start\n        start = 0\n\n    if end > 60:\n        start = start - (end-60)\n        end = 60\n\n    start_idx = sec2index(start, spec_len)\n    end_idx = sec2index(end, spec_len)\n\n    return start_idx, end_idx\n\nmel_spec = librosa.feature.melspectrogram(audio_vec, n_fft=FFT, hop_length=HOP, sr=SR, fmin=F_MIN, fmax=F_MAX, power=POWER)\n\nspec_len = mel_spec.shape[1]\n\nstart, end = make_slice(t_mid, WINDOW_WIDTH, spec_len)\n\nmel_spec_slice = mel_spec[:, start:end]\n
\n

I'm sure you can modify this code to work on the y-axis/frequency axis as well.

", "votes": null, "replies": [] }, { "id": 1156822, "author_name": "aryaman1999", "author_url": "", "post_date": "01/17/2021 12:37:43", "content": "

Thanks for comment @RNA, but for some reason the same logic doesnt really translate to the y axis :/

", "votes": null, "replies": [] }, { "id": 1156939, "author_name": "aryaman1999", "author_url": "", "post_date": "01/17/2021 14:38:01", "content": "

I solved it by limiting the librosa mel filter bank to the adequate freq limits, in case any one runs into the same problem

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1154573": "After studying the data a bit I realised that a single audio file has multiple bird sounds as pointed out by @kneroma .\nBut I am still confused as to how does one crop audio? Should I create an image of all the spectrograms and then use the image co-ordinates to make the crops? Can I do it directly on the spectrogram without converting the spectrogram to an image? If yes then how?\n\nHeres the bounding boxes I want to crop.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1604765%2F04415575fe06bdcc94041eed48acdf1f%2FScreenshot%20from%202021-01-15%2023-26-07.png?generation=1610733598342597&alt=media)", "1154651": "The spectogram is a 2D numpy array that is easily subscriptable. For example, my code to isolate a section with a midpoint (in seconds) and a fixed width:\n\n```\ndef sec2index(s, spec_len=5626):\n return int(np.round(s*spec_len/60))\n\ndef make_slice(mid, width, spec_len):\n end = mid + width\n start = mid - width\n \n if start < 0:\n end = end - start\n start = 0\n \n if end > 60:\n start = start - (end-60)\n end = 60\n \n start_idx = sec2index(start, spec_len)\n end_idx = sec2index(end, spec_len)\n \n return start_idx, end_idx\n\nmel_spec = librosa.feature.melspectrogram(audio_vec, n_fft=FFT, hop_length=HOP, sr=SR, fmin=F_MIN, fmax=F_MAX, power=POWER)\n\nspec_len = mel_spec.shape[1]\n\nstart, end = make_slice(t_mid, WINDOW_WIDTH, spec_len)\n\nmel_spec_slice = mel_spec[:, start:end]\n```\n\nI'm sure you can modify this code to work on the y-axis/frequency axis as well.", "1156822": "Thanks for comment @RNA, but for some reason the same logic doesnt really translate to the y axis :/", "1156939": "I solved it by limiting the librosa mel filter bank to the adequate freq limits, in case any one runs into the same problem" }, "source": "meta" }