Is it related to the number of Mel-filterbanks to extract features?
\nI have used the value: 128
\nI wonder if the value improves the model's accuracy.
", "messages": [ { "id": "1189722", "postDate": "02/07/2021 07:53:47", "content": "Is it related to the number of Mel-filterbanks to extract features?
\nI have used the value: 128
\nI wonder if the value improves the model's accuracy.
", "rawMarkdown": "Is it related to the number of Mel-filterbanks to extract features?\n\nI have used the value: 128\n\nI wonder if the value improves the model's accuracy.", "votes": null }, { "id": "1189817", "postDate": "02/07/2021 08:55:35", "content": "I'm using 64, because 128 gives me \"empty channels\" warning.
", "rawMarkdown": "I'm using 64, because 128 gives me \"empty channels\" warning.", "votes": null }, { "id": "1190156", "postDate": "02/07/2021 14:17:46", "content": "Increasing the number of mel filterbanks helps, but your training time will also increase.
", "rawMarkdown": "Increasing the number of mel filterbanks helps, but your training time will also increase.", "votes": null }, { "id": "1190158", "postDate": "02/07/2021 14:21:29", "content": "In this case, you have to increase n_fft.
", "rawMarkdown": "In this case, you have to increase n_fft.", "votes": null }, { "id": "1190786", "postDate": "02/08/2021 02:16:55", "content": "Thank you for your reply.
\nLooking at your answer to the question below, I would like to ask if I should lower the n_fft if I raise the value of the number of Mel filterbanks.
", "rawMarkdown": "Thank you for your reply. \n\nLooking at your answer to the question below, I would like to ask if I should lower the n_fft if I raise the value of the number of Mel filterbanks.", "votes": null }, { "id": "1190903", "postDate": "02/08/2021 05:17:48", "content": "My understanding (and I'm not an expert on spectrograms), is that making a mel spectrogram on a 60 second sound file creates a 2D array of height H and width W.
Increasing mel filters increases the height H (the frequency dimension), and you should set n_fft large enough so that you don't get \"empty channels\". (Basically n_fft and mel filters are the same thing. If you do a basic spectrogram then n_fft is your height. Next choosing mel filters picks a new height and just rescales the y axis when converting a spectrogram into a mel spectrogram)
Next you can consider adjusting the width W (the time dimension) by adjusting hop_length and win_length. Note that there is a tradeoff. You can accomplish any combination of H and W you desire. But your selection of hop_length and win_length decides whether there is more information in frequency H or time W.
For example, just because you have a large H or W doesn't mean you have good resolution (i.e. lots of information in that dimension). Depending on hop_length and win_length, either the frequency or time is \"blurred\" and doesn't contain full information in that dimension.
\"Looking at your answer to the question below, I would like to ask if I should lower the n_fft if I raise the value of the number of Mel filterbanks.\"
\nLike Chris has said, they have a direct relationship, not an inverse relation. Please don't lower n_fft if you increase mel filter. I believe with default n_fft of 2048, you should be able to increase the mels from 128 to 256 without any issue and see a slight improvement in score. You can refer to : https://www.kaggle.com/allohvk/the-nuts-and-bolts-of-sound-feature-engineering for a theoretical explanation of these params and a peek into what they actually signify
", "rawMarkdown": "*\"Looking at your answer to the question below, I would like to ask if I should lower the n_fft if I raise the value of the number of Mel filterbanks.\"*\n\nLike Chris has said, they have a direct relationship, not an inverse relation. Please don't lower n_fft if you increase mel filter. I believe with default n_fft of 2048, you should be able to increase the mels from 128 to 256 without any issue and see a slight improvement in score. You can refer to : https://www.kaggle.com/allohvk/the-nuts-and-bolts-of-sound-feature-engineering for a theoretical explanation of these params and a peek into what they actually signify", "votes": null }, { "id": "1191199", "postDate": "02/08/2021 10:14:40", "content": "@cdeotte @allohvk Thanks for the detailed explanations.
\nThen I will try to the values and consider another model to improve the score.
Thank you for your reply.
", "rawMarkdown": "Thank you for your reply.", "votes": null }, { "id": "2945678", "postDate": "08/03/2024 16:31:36", "content": "Very interesting
", "rawMarkdown": "Very interesting", "votes": null } ], "comments": [ { "id": 1189817, "author_name": "atamazian", "author_url": "", "post_date": "02/07/2021 08:55:35", "content": "I'm using 64, because 128 gives me \"empty channels\" warning.
", "votes": null, "replies": [ { "id": 1190158, "author_name": "tugstugi", "author_url": "", "post_date": "02/07/2021 14:21:29", "content": "In this case, you have to increase n_fft.
", "votes": null, "replies": [] }, { "id": 1191200, "author_name": "kbh0287", "author_url": "", "post_date": "02/08/2021 10:16:42", "content": "Thank you for your reply.
", "votes": null, "replies": [] } ] }, { "id": 1190156, "author_name": "tugstugi", "author_url": "", "post_date": "02/07/2021 14:17:46", "content": "Increasing the number of mel filterbanks helps, but your training time will also increase.
", "votes": null, "replies": [ { "id": 1190786, "author_name": "kbh0287", "author_url": "", "post_date": "02/08/2021 02:16:55", "content": "Thank you for your reply.
\nLooking at your answer to the question below, I would like to ask if I should lower the n_fft if I raise the value of the number of Mel filterbanks.
", "votes": null, "replies": [] }, { "id": 1190903, "author_name": "cdeotte", "author_url": "", "post_date": "02/08/2021 05:17:48", "content": "My understanding (and I'm not an expert on spectrograms), is that making a mel spectrogram on a 60 second sound file creates a 2D array of height H and width W.
Increasing mel filters increases the height H (the frequency dimension), and you should set n_fft large enough so that you don't get \"empty channels\". (Basically n_fft and mel filters are the same thing. If you do a basic spectrogram then n_fft is your height. Next choosing mel filters picks a new height and just rescales the y axis when converting a spectrogram into a mel spectrogram)
Next you can consider adjusting the width W (the time dimension) by adjusting hop_length and win_length. Note that there is a tradeoff. You can accomplish any combination of H and W you desire. But your selection of hop_length and win_length decides whether there is more information in frequency H or time W.
For example, just because you have a large H or W doesn't mean you have good resolution (i.e. lots of information in that dimension). Depending on hop_length and win_length, either the frequency or time is \"blurred\" and doesn't contain full information in that dimension.
\"Looking at your answer to the question below, I would like to ask if I should lower the n_fft if I raise the value of the number of Mel filterbanks.\"
\nLike Chris has said, they have a direct relationship, not an inverse relation. Please don't lower n_fft if you increase mel filter. I believe with default n_fft of 2048, you should be able to increase the mels from 128 to 256 without any issue and see a slight improvement in score. You can refer to : https://www.kaggle.com/allohvk/the-nuts-and-bolts-of-sound-feature-engineering for a theoretical explanation of these params and a peek into what they actually signify
", "votes": null, "replies": [] }, { "id": 1191199, "author_name": "kbh0287", "author_url": "", "post_date": "02/08/2021 10:14:40", "content": "@cdeotte @allohvk Thanks for the detailed explanations.
\nThen I will try to the values and consider another model to improve the score.
Very interesting
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1189722": "Is it related to the number of Mel-filterbanks to extract features?\n\nI have used the value: 128\n\nI wonder if the value improves the model's accuracy.", "1189817": "I'm using 64, because 128 gives me \"empty channels\" warning.", "1190156": "Increasing the number of mel filterbanks helps, but your training time will also increase.", "1190158": "In this case, you have to increase n_fft.", "1190786": "Thank you for your reply. \n\nLooking at your answer to the question below, I would like to ask if I should lower the n_fft if I raise the value of the number of Mel filterbanks.", "1190903": "My understanding (and I'm not an expert on spectrograms), is that making a mel spectrogram on a 60 second sound file creates a 2D array of height `H` and width `W`.\n\nIncreasing `mel filters` increases the height `H` (the frequency dimension), and you should set `n_fft` large enough so that you don't get \"empty channels\". (Basically `n_fft` and `mel filters` are the same thing. If you do a basic spectrogram then `n_fft` is your height. Next choosing `mel filters` picks a new height and just rescales the y axis when converting a spectrogram into a mel spectrogram) \n\nNext you can consider adjusting the width `W` (the time dimension) by adjusting `hop_length` and `win_length`. Note that there is a tradeoff. You can accomplish any combination of `H` and `W` you desire. But your selection of `hop_length` and `win_length` decides whether there is more information in frequency `H` or time `W`.\n\nFor example, just because you have a large `H` or `W` doesn't mean you have good resolution (i.e. lots of information in that dimension). Depending on `hop_length` and `win_length`, either the frequency or time is \"blurred\" and doesn't contain full information in that dimension.", "1191063": "*\"Looking at your answer to the question below, I would like to ask if I should lower the n_fft if I raise the value of the number of Mel filterbanks.\"*\n\nLike Chris has said, they have a direct relationship, not an inverse relation. Please don't lower n_fft if you increase mel filter. I believe with default n_fft of 2048, you should be able to increase the mels from 128 to 256 without any issue and see a slight improvement in score. You can refer to : https://www.kaggle.com/allohvk/the-nuts-and-bolts-of-sound-feature-engineering for a theoretical explanation of these params and a peek into what they actually signify", "1191199": "cdeotte @allohvk Thanks for the detailed explanations. \nThen I will try to the values and consider another model to improve the score.", "1191200": "Thank you for your reply.", "2945678": "Very interesting" }, "source": "meta" }