During the search of toolboxes I've stumbled upon paper which provides quite good comparison of famous instruments to use. Hope it will be helpful for you): https://www.ntnu.edu/documents/1001201110/1266017954/DAFx-15_submission_43_v2.pdf/06508f48-9272-41c8-9381-7639a0240770\nP.s: On \"Essentia\" web site there are, also, many useful papers on related topics.
", "messages": [ { "id": "348931", "postDate": "06/27/2018 14:44:29", "content": "During the search of toolboxes I've stumbled upon paper which provides quite good comparison of famous instruments to use. Hope it will be helpful for you): https://www.ntnu.edu/documents/1001201110/1266017954/DAFx-15_submission_43_v2.pdf/06508f48-9272-41c8-9381-7639a0240770\nP.s: On \"Essentia\" web site there are, also, many useful papers on related topics.
", "rawMarkdown": "During the search of toolboxes I've stumbled upon paper which provides quite good comparison of famous instruments to use. Hope it will be helpful for you): https://www.ntnu.edu/documents/1001201110/1266017954/DAFx-15_submission_43_v2.pdf/06508f48-9272-41c8-9381-7639a0240770\nP.s: On \"Essentia\" web site there are, also, many useful papers on related topics.", "votes": null }, { "id": "350198", "postDate": "06/29/2018 11:38:45", "content": "Thanks for sharing, Essentia seems to be good!
", "rawMarkdown": "Thanks for sharing, Essentia seems to be good!", "votes": null }, { "id": "350716", "postDate": "06/30/2018 09:51:10", "content": "No problem) I've choose Essentia, too. Freesound actually also use it for large-scale indexing and content-based search of sound recordings: http://essentia.upf.edu/documentation/applications.html
", "rawMarkdown": "No problem) I've choose Essentia, too. Freesound actually also use it for large-scale indexing and content-based search of sound recordings: http://essentia.upf.edu/documentation/applications.html", "votes": null }, { "id": "350868", "postDate": "06/30/2018 16:31:58", "content": "Thanks again, that sounds kind of important information for this competition; it seems to be hard to get better score unless tracking how data was labeled. :) Anyway I'd like to start looking into Essentia.
", "rawMarkdown": "Thanks again, that sounds kind of important information for this competition; it seems to be hard to get better score unless tracking how data was labeled. :) Anyway I'd like to start looking into Essentia.", "votes": null }, { "id": "351473", "postDate": "07/02/2018 09:13:29", "content": "Just my 2 cents – YAAFE is really fast, and the syntax, while being not that nice in Python (basically just passing a lot of strings describing what features you want to extract), is not that much of a problem.\nSo for extracting basic things like MFCC / spectral shape statistics / etc., YAAFE works really well.
", "rawMarkdown": "Just my 2 cents – YAAFE is really fast, and the syntax, while being not that nice in Python (basically just passing a lot of strings describing what features you want to extract), is not that much of a problem.\nSo for extracting basic things like MFCC / spectral shape statistics / etc., YAAFE works really well.", "votes": null }, { "id": "352487", "postDate": "07/04/2018 13:16:21", "content": "For those interested in using Essentia, you can now install it for Python on Linux using \"pip install essentia\".
\n\nYou can use the FreesoundExtractor or MusicExtractor algorithms in Python to extract lots of frame-level features and their statistical summarization for any given audio file.
", "rawMarkdown": "For those interested in using Essentia, you can now install it for Python on Linux using \"pip install essentia\". \n\nYou can use the [FreesoundExtractor][1] or [MusicExtractor][2] algorithms in Python to extract lots of frame-level features and their statistical summarization for any given audio file.\n\n\n [1]: http://essentia.upf.edu/documentation/reference/std_FreesoundExtractor.html\n [2]: http://essentia.upf.edu/documentation/reference/std_MusicExtractor.html", "votes": null }, { "id": "359319", "postDate": "07/19/2018 21:32:49", "content": "I prefer to do all the audio feature extractions/transformations using just plain NumPy (and SciPy for reading wavs). That way, I have better intuition and control over the parameters than if I just used libraries. Even though the performance might suffer a little compared to C/C++ library implementations (have not tested though). Also, what I do and recommend is to first cache the spectrograms, wavelets, features ... using either pickle or h5py before you start training. If you do that first, you can then load all data (ready for training) into memory in less than 20 seconds on a 6-core CPU and you can test/iterate multiple models without the overhead of recreating the same features.
", "rawMarkdown": "I prefer to do all the audio feature extractions/transformations using just plain NumPy (and SciPy for reading wavs). That way, I have better intuition and control over the parameters than if I just used libraries. Even though the performance might suffer a little compared to C/C++ library implementations (have not tested though). Also, what I do and recommend is to first cache the spectrograms, wavelets, features ... using either pickle or h5py before you start training. If you do that first, you can then load all data (ready for training) into memory in less than 20 seconds on a 6-core CPU and you can test/iterate multiple models without the overhead of recreating the same features.", "votes": null } ], "comments": [ { "id": 350198, "author_name": "daisukelab", "author_url": "", "post_date": "06/29/2018 11:38:45", "content": "Thanks for sharing, Essentia seems to be good!
", "votes": null, "replies": [ { "id": 350716, "author_name": "stasbuzuluk", "author_url": "", "post_date": "06/30/2018 09:51:10", "content": "No problem) I've choose Essentia, too. Freesound actually also use it for large-scale indexing and content-based search of sound recordings: http://essentia.upf.edu/documentation/applications.html
", "votes": null, "replies": [] }, { "id": 350868, "author_name": "daisukelab", "author_url": "", "post_date": "06/30/2018 16:31:58", "content": "Thanks again, that sounds kind of important information for this competition; it seems to be hard to get better score unless tracking how data was labeled. :) Anyway I'd like to start looking into Essentia.
", "votes": null, "replies": [] } ] }, { "id": 351473, "author_name": "kunstmord", "author_url": "", "post_date": "07/02/2018 09:13:29", "content": "Just my 2 cents – YAAFE is really fast, and the syntax, while being not that nice in Python (basically just passing a lot of strings describing what features you want to extract), is not that much of a problem.\nSo for extracting basic things like MFCC / spectral shape statistics / etc., YAAFE works really well.
", "votes": null, "replies": [] }, { "id": 352487, "author_name": "dibogdanov", "author_url": "", "post_date": "07/04/2018 13:16:21", "content": "For those interested in using Essentia, you can now install it for Python on Linux using \"pip install essentia\".
\n\nYou can use the FreesoundExtractor or MusicExtractor algorithms in Python to extract lots of frame-level features and their statistical summarization for any given audio file.
", "votes": null, "replies": [] }, { "id": 359319, "author_name": "danielhavir", "author_url": "", "post_date": "07/19/2018 21:32:49", "content": "I prefer to do all the audio feature extractions/transformations using just plain NumPy (and SciPy for reading wavs). That way, I have better intuition and control over the parameters than if I just used libraries. Even though the performance might suffer a little compared to C/C++ library implementations (have not tested though). Also, what I do and recommend is to first cache the spectrograms, wavelets, features ... using either pickle or h5py before you start training. If you do that first, you can then load all data (ready for training) into memory in less than 20 seconds on a 6-core CPU and you can test/iterate multiple models without the overhead of recreating the same features.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "348931": "During the search of toolboxes I've stumbled upon paper which provides quite good comparison of famous instruments to use. Hope it will be helpful for you): https://www.ntnu.edu/documents/1001201110/1266017954/DAFx-15_submission_43_v2.pdf/06508f48-9272-41c8-9381-7639a0240770\nP.s: On \"Essentia\" web site there are, also, many useful papers on related topics.", "350198": "Thanks for sharing, Essentia seems to be good!", "350716": "No problem) I've choose Essentia, too. Freesound actually also use it for large-scale indexing and content-based search of sound recordings: http://essentia.upf.edu/documentation/applications.html", "350868": "Thanks again, that sounds kind of important information for this competition; it seems to be hard to get better score unless tracking how data was labeled. :) Anyway I'd like to start looking into Essentia.", "351473": "Just my 2 cents – YAAFE is really fast, and the syntax, while being not that nice in Python (basically just passing a lot of strings describing what features you want to extract), is not that much of a problem.\nSo for extracting basic things like MFCC / spectral shape statistics / etc., YAAFE works really well.", "352487": "For those interested in using Essentia, you can now install it for Python on Linux using \"pip install essentia\". \n\nYou can use the [FreesoundExtractor][1] or [MusicExtractor][2] algorithms in Python to extract lots of frame-level features and their statistical summarization for any given audio file.\n\n\n [1]: http://essentia.upf.edu/documentation/reference/std_FreesoundExtractor.html\n [2]: http://essentia.upf.edu/documentation/reference/std_MusicExtractor.html", "359319": "I prefer to do all the audio feature extractions/transformations using just plain NumPy (and SciPy for reading wavs). That way, I have better intuition and control over the parameters than if I just used libraries. Even though the performance might suffer a little compared to C/C++ library implementations (have not tested though). Also, what I do and recommend is to first cache the spectrograms, wavelets, features ... using either pickle or h5py before you start training. If you do that first, you can then load all data (ready for training) into memory in less than 20 seconds on a 6-core CPU and you can test/iterate multiple models without the overhead of recreating the same features." }, "source": "meta" }