{ "id": 212638, "title": "Best practices for creating spectrograms?", "url": "/competitions/rfcx-species-audio-detection/discussion/212638", "author_name": "", "post_date": "2021-01-19T17:05:51.910525900Z", "votes": 1, "comment_count": 2, "views": 0, "content": "

What spectrogram parameters and augmentations have you guys found to work best?

\n

Currently I've been using using:
\n# create Mel spectrogram
\nstft = librosa.feature.melspectrogram(loaded_audio, sr=sample_rate, power=power,\n fmin=f_min, fmax=f_max, n_mels=shape[0])

\n

stft_to_db = librosa.core.amplitude_to_db(np.abs(stft))

\n

With

\n

shape = (256,512)
\nsample_rate = 48000 which is the native sampling rate of the audio file
\npower=1.5
\nf_min = round(df_in.f_min.min() * .75, 1) -> min freq reduced by 25%
\nf_max = round(df_in.f_max.max() * 1.25, 1) -> max freq increased by 25%
\nn_mels = 256 (x length of image size)

\n

Augmentations:

\n\n

Example Spectrogram:
\n\"\"

\n

Full code: https://www.kaggle.com/stegzz/readable-keras-resnet50-model

", "messages": [ { "id": "1160091", "postDate": "01/19/2021 17:05:51", "content": "

What spectrogram parameters and augmentations have you guys found to work best?

\n

Currently I've been using using:
\n# create Mel spectrogram
\nstft = librosa.feature.melspectrogram(loaded_audio, sr=sample_rate, power=power,\n fmin=f_min, fmax=f_max, n_mels=shape[0])

\n

stft_to_db = librosa.core.amplitude_to_db(np.abs(stft))

\n

With

\n

shape = (256,512)
\nsample_rate = 48000 which is the native sampling rate of the audio file
\npower=1.5
\nf_min = round(df_in.f_min.min() * .75, 1) -> min freq reduced by 25%
\nf_max = round(df_in.f_max.max() * 1.25, 1) -> max freq increased by 25%
\nn_mels = 256 (x length of image size)

\n

Augmentations:

\n\n

Example Spectrogram:
\n\"\"

\n

Full code: https://www.kaggle.com/stegzz/readable-keras-resnet50-model

", "rawMarkdown": "What spectrogram parameters and augmentations have you guys found to work best?\n\nCurrently I've been using using:\n`# create Mel spectrogram`\n`stft = librosa.feature.melspectrogram(loaded_audio, sr=sample_rate, power=power,\n fmin=f_min, fmax=f_max, n_mels=shape[0])`\n\n`stft_to_db = librosa.core.amplitude_to_db(np.abs(stft))`\n\n\nWith \n\n`shape = (256,512)` \n`sample_rate = 48000` which is the native sampling rate of the audio file\n`power=1.5`\n`f_min = round(df_in.f_min.min() * .75, 1)` -> min freq reduced by 25%\n`f_max = round(df_in.f_max.max() * 1.25, 1)` -> max freq increased by 25%\n`n_mels = 256` (x length of image size)\n\nAugmentations:\n- Random offset of event plus or minus 3 seconds\n- Adding random gaussian noise to spectrogram\n- Random time masking\n- Random frequency masking\n\nExample Spectrogram:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3436076%2Ffbe3decc95af953c4073fff60cec1b7b%2Ftest.png?generation=1611075880155421&alt=media)\n\nFull code: https://www.kaggle.com/stegzz/readable-keras-resnet50-model", "votes": null }, { "id": "1160140", "postDate": "01/19/2021 17:31:53", "content": "

Have you tried adding mixup?

", "rawMarkdown": "Have you tried adding mixup?", "votes": null }, { "id": "1160178", "postDate": "01/19/2021 18:06:32", "content": "

There is only one way to know: try various ways and see which settings lead to best model (better CV and LB).

", "rawMarkdown": "There is only one way to know: try various ways and see which settings lead to best model (better CV and LB).", "votes": null } ], "comments": [ { "id": 1160140, "author_name": "atamazian", "author_url": "", "post_date": "01/19/2021 17:31:53", "content": "

Have you tried adding mixup?

", "votes": null, "replies": [] }, { "id": 1160178, "author_name": "cpmpml", "author_url": "", "post_date": "01/19/2021 18:06:32", "content": "

There is only one way to know: try various ways and see which settings lead to best model (better CV and LB).

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1160091": "What spectrogram parameters and augmentations have you guys found to work best?\n\nCurrently I've been using using:\n`# create Mel spectrogram`\n`stft = librosa.feature.melspectrogram(loaded_audio, sr=sample_rate, power=power,\n fmin=f_min, fmax=f_max, n_mels=shape[0])`\n\n`stft_to_db = librosa.core.amplitude_to_db(np.abs(stft))`\n\n\nWith \n\n`shape = (256,512)` \n`sample_rate = 48000` which is the native sampling rate of the audio file\n`power=1.5`\n`f_min = round(df_in.f_min.min() * .75, 1)` -> min freq reduced by 25%\n`f_max = round(df_in.f_max.max() * 1.25, 1)` -> max freq increased by 25%\n`n_mels = 256` (x length of image size)\n\nAugmentations:\n- Random offset of event plus or minus 3 seconds\n- Adding random gaussian noise to spectrogram\n- Random time masking\n- Random frequency masking\n\nExample Spectrogram:\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3436076%2Ffbe3decc95af953c4073fff60cec1b7b%2Ftest.png?generation=1611075880155421&alt=media)\n\nFull code: https://www.kaggle.com/stegzz/readable-keras-resnet50-model", "1160140": "Have you tried adding mixup?", "1160178": "There is only one way to know: try various ways and see which settings lead to best model (better CV and LB)." }, "source": "meta" }