![\"image\"](\"https://user-images.githubusercontent.com/58644957/154795828-9bda47c1-bf27-4388-b114-4d45a0d6429a.png\")
I found that some sound files had (X, 2) dim.
\nI'm not familiar with it, but is the difference due to the author's recording equipment?
\nAlso, is averaging a good way to handle this kind of audio?
data = np.mean(data, 1)
Thanks.
", "messages": [ { "id": "1697007", "postDate": "02/19/2022 09:54:03", "content": "I found that some sound files had (X, 2) dim.
\nI'm not familiar with it, but is the difference due to the author's recording equipment?
\nAlso, is averaging a good way to handle this kind of audio?
data = np.mean(data, 1)
Thanks.
", "rawMarkdown": "I found that some sound files had (X, 2) dim.\nI'm not familiar with it, but is the difference due to the author's recording equipment?\nAlso, is averaging a good way to handle this kind of audio?\n\n`data = np.mean(data, 1)`\n\n\n\nThanks.", "votes": null }, { "id": "1697035", "postDate": "02/19/2022 10:18:20", "content": "That means binaural. You can use librosa.to_mono to uniformly convert to mono format.
@zzy990106 Thanks for sharing the function, I wasn't aware of it!
\nDon't you think we are discarding useful info by converting binaural to mono format?
\nI picture it's like converting an RGB image to B/W, throwing away info that models could potentially use.
", "rawMarkdown": "zzy990106 Thanks for sharing the function, I wasn't aware of it! \n\nDon't you think we are discarding useful info by converting binaural to mono format? \n\nI picture it's like converting an RGB image to B/W, throwing away info that models could potentially use.", "votes": null }, { "id": "1697071", "postDate": "02/19/2022 10:48:55", "content": "Yes, there is information loss, but the impact on classification is far negligible.
", "rawMarkdown": "Yes, there is information loss, but the impact on classification is far negligible.", "votes": null }, { "id": "1697388", "postDate": "02/19/2022 15:38:15", "content": "@zzy990106, is there a difference between binaural and stereo sound in this case ?
\nCan't we just take the mean of the 2 channels ?
\ny, sr = sf.read(path, always_2d=True)
y = np.mean(y, 1)
Alex, I believe both are the same
\n\n\nCan't we just take the mean of the 2 channels ?
\n
I think this might be the best approach! I'll try this 👌. Thanks!
", "rawMarkdown": "Alex, I believe both are the same\n\n> Can't we just take the mean of the 2 channels ?\n\nI think this might be the best approach! I'll try this 👌. Thanks!", "votes": null }, { "id": "1697899", "postDate": "02/20/2022 00:28:04", "content": "@zzy990106, Thanks for the answer!
", "rawMarkdown": "zzy990106, Thanks for the answer!", "votes": null }, { "id": "1699305", "postDate": "02/21/2022 05:14:02", "content": "I have experimented with 3 approaches to this problem.
\nI thought that approach 3 would work, but the results of LB were as follows
\n| approach | \nLB | \n
|---|---|
| 1 | \n0.70 | \n
| 2 | \n0.69 | \n
| 3 | \n0.64 | \n
For reference, as you know, all test audio files seem to have only one channel. I don't know if this was recorded monaural, or if they do some way to make it only one channel…
", "rawMarkdown": "I have experimented with 3 approaches to this problem.\n\n1. use only the first channel.\n2. calculate the mean.\n3. randomly sample at training which channel to use.\n\nI thought that approach 3 would work, but the results of LB were as follows\n\n| approach | LB |\n| --- | --- |\n| 1 | 0.70 |\n| 2 | 0.69 |\n| 3 | 0.64 |\n\nFor reference, as you know, all test audio files seem to have only one channel. I don't know if this was recorded monaural, or if they do some way to make it only one channel...", "votes": null }, { "id": "1699522", "postDate": "02/21/2022 08:41:28", "content": "Could you share how you only use the first channel ?
", "rawMarkdown": "Could you share how you only use the first channel ?", "votes": null }, { "id": "1699582", "postDate": "02/21/2022 09:41:13", "content": "This is an example of assigning the first channel to y.
\ny, sr = sf.read(wav_path, always_2d=True)\ny = y[:, 0] \nI found there are some date that the volume of one microphone is abnormal (for example, XC344134.ogg : Channel 1 broken ,XC521984.ogg : Channel 0 broken).
\nIf you normalize the data after transform to spectrogram, it might not be a problem, but be careful when dealing with volumes.
@nomorevotch I had same idea as you (randomly select one of the audio channel when training) and my CV is quite lower than when I use channel average. Kind of confirm your experience.
", "rawMarkdown": "nomorevotch I had same idea as you (randomly select one of the audio channel when training) and my CV is quite lower than when I use channel average. Kind of confirm your experience.", "votes": null } ], "comments": [ { "id": 1697035, "author_name": "zzy990106", "author_url": "", "post_date": "02/19/2022 10:18:20", "content": "That means binaural. You can use librosa.to_mono to uniformly convert to mono format.
@zzy990106 Thanks for sharing the function, I wasn't aware of it!
\nDon't you think we are discarding useful info by converting binaural to mono format?
\nI picture it's like converting an RGB image to B/W, throwing away info that models could potentially use.
", "votes": null, "replies": [] }, { "id": 1697071, "author_name": "zzy990106", "author_url": "", "post_date": "02/19/2022 10:48:55", "content": "Yes, there is information loss, but the impact on classification is far negligible.
", "votes": null, "replies": [] }, { "id": 1697388, "author_name": "alexandrecc", "author_url": "", "post_date": "02/19/2022 15:38:15", "content": "@zzy990106, is there a difference between binaural and stereo sound in this case ?
\nCan't we just take the mean of the 2 channels ?
\ny, sr = sf.read(path, always_2d=True)
y = np.mean(y, 1)
Alex, I believe both are the same
\n\n\nCan't we just take the mean of the 2 channels ?
\n
I think this might be the best approach! I'll try this 👌. Thanks!
", "votes": null, "replies": [] }, { "id": 1697899, "author_name": "naoism", "author_url": "", "post_date": "02/20/2022 00:28:04", "content": "@zzy990106, Thanks for the answer!
", "votes": null, "replies": [] }, { "id": 1699305, "author_name": "nomorevotch", "author_url": "", "post_date": "02/21/2022 05:14:02", "content": "I have experimented with 3 approaches to this problem.
\nI thought that approach 3 would work, but the results of LB were as follows
\n| approach | \nLB | \n
|---|---|
| 1 | \n0.70 | \n
| 2 | \n0.69 | \n
| 3 | \n0.64 | \n
For reference, as you know, all test audio files seem to have only one channel. I don't know if this was recorded monaural, or if they do some way to make it only one channel…
", "votes": null, "replies": [] }, { "id": 1699522, "author_name": "mithilsalunkhe", "author_url": "", "post_date": "02/21/2022 08:41:28", "content": "Could you share how you only use the first channel ?
", "votes": null, "replies": [] }, { "id": 1699582, "author_name": "nomorevotch", "author_url": "", "post_date": "02/21/2022 09:41:13", "content": "This is an example of assigning the first channel to y.
\ny, sr = sf.read(wav_path, always_2d=True)\ny = y[:, 0] \n@nomorevotch I had same idea as you (randomly select one of the audio channel when training) and my CV is quite lower than when I use channel average. Kind of confirm your experience.
", "votes": null, "replies": [] } ] }, { "id": 1701393, "author_name": "asaliquid1011", "author_url": "", "post_date": "02/22/2022 18:10:03", "content": "I found there are some date that the volume of one microphone is abnormal (for example, XC344134.ogg : Channel 1 broken ,XC521984.ogg : Channel 0 broken).
\nIf you normalize the data after transform to spectrogram, it might not be a problem, but be careful when dealing with volumes.