{ "id": 220315, "title": "My First Competition Medal (Top %7)", "url": "/competitions/rfcx-species-audio-detection/discussion/220315", "author_name": "", "post_date": "2021-02-18T00:46:46.749928Z", "votes": 31, "comment_count": 17, "views": 0, "content": "

Hi everyone,

\n

First of all, congrats to all the winners, and thanks to the host for this interesting challenge.

\n

I'm very happy that I ended up in the same position in the private LB as the public and won a bronze medal 😊😊😊. This was kinda the first competition that I heavily spend time on. So, it means a lot to me 😊

\n

I want to thanks @barnwellguy @hidehisaarai1213 @gopidurgaprasad @cdeotte for their insightful comments and for sharing great ideas.

\n

Here is the summary/story of my solution for the problem:

\n

This was my first audio competition so I had to do some research before doing any experimentation. I started watching some videos from this video series. After grasping some intuition about the problem, I did some data visualization and tried to understand how the sounds change with different classes.

\n

For the modeling process, I used this notebook as my baseline and produced my initial models. I had a ~0.75 score from the initial experimentations. Then I started looking at old audio competitions and wanted to see what other people did. So, I did some research on Cornell competition. In that competition, I saw the @hidehisaarai1213 's solution. The code was available on GitHub and it was very well modularized and documented. I mostly focused on the feature representation part from his code. By using the pcen additional to melspectrogram for the feature representation, I had a ~0.81 score. Then, I wanted to understand how the data augmentation is implemented for the audio data so I opened this discussion in the forums. With the @barnwellguy 's help out there, I managed to add the data augmentation to my pipeline and achieved a ~0.84 score. That was the first part of my solution. I was using resnest50 in these experimentations.

\n

Then, I wanted to learn about different modeling strategies and researched a bit more on that part. In the Cornell competition, there were many solutions which were including a SED model. And, there is a great introductory notebook which was written by @hidehisaarai1213 again πŸ™ I carefully studied this notebook, I'm not going to lie, it was kinda overwhelming at first glance, but I believe that it was worth it. There was also an introductory notebook which was based on @hidehisaarai1213 's notebook written by @gopidurgaprasad in this competition. By playing with the different parameters (I was much comfortable after studying the original notebook), I managed to achieve a ~0.83 score. After reading a comment from @cdeotte 's (couldn't find the link), I played with the input parameters again (like n_mel, hop_length) and increased my score to ~0.85. I was using effnet_b0 for all of my experiments until then. I switched to b2, b3, b4 and had a score improvement of 0.863, 0.873, 0.88 respectively.

\n

Then, I did rank-based bagging of different models produced along the way and achieved a 0.90 score in the public LB. Finally, inspired by this discussion I used rankdata from the scipy for the ranking part and improved my score to 0.902 for my final submission in the public LB.

\n

Here's also my parameters:

\n

For the SED models:

\n\n

For the standart image models:

\n\n

I used only TP data for training. Tried to add FP data as well but couldn't manage it to work, unfortunately.

\n

Thanks everyone again for this competition journey. I learned a lot along the way.

\n

Edit:

\n

My models private LB was 0.90990.

\n

After applying @cdeotte 's clever postprocessing method it jumps to 0.94729

\n", "messages": [ { "id": "1207659", "postDate": "02/18/2021 00:46:46", "content": "

Hi everyone,

\n

First of all, congrats to all the winners, and thanks to the host for this interesting challenge.

\n

I'm very happy that I ended up in the same position in the private LB as the public and won a bronze medal 😊😊😊. This was kinda the first competition that I heavily spend time on. So, it means a lot to me 😊

\n

I want to thanks @barnwellguy @hidehisaarai1213 @gopidurgaprasad @cdeotte for their insightful comments and for sharing great ideas.

\n

Here is the summary/story of my solution for the problem:

\n

This was my first audio competition so I had to do some research before doing any experimentation. I started watching some videos from this video series. After grasping some intuition about the problem, I did some data visualization and tried to understand how the sounds change with different classes.

\n

For the modeling process, I used this notebook as my baseline and produced my initial models. I had a ~0.75 score from the initial experimentations. Then I started looking at old audio competitions and wanted to see what other people did. So, I did some research on Cornell competition. In that competition, I saw the @hidehisaarai1213 's solution. The code was available on GitHub and it was very well modularized and documented. I mostly focused on the feature representation part from his code. By using the pcen additional to melspectrogram for the feature representation, I had a ~0.81 score. Then, I wanted to understand how the data augmentation is implemented for the audio data so I opened this discussion in the forums. With the @barnwellguy 's help out there, I managed to add the data augmentation to my pipeline and achieved a ~0.84 score. That was the first part of my solution. I was using resnest50 in these experimentations.

\n

Then, I wanted to learn about different modeling strategies and researched a bit more on that part. In the Cornell competition, there were many solutions which were including a SED model. And, there is a great introductory notebook which was written by @hidehisaarai1213 again πŸ™ I carefully studied this notebook, I'm not going to lie, it was kinda overwhelming at first glance, but I believe that it was worth it. There was also an introductory notebook which was based on @hidehisaarai1213 's notebook written by @gopidurgaprasad in this competition. By playing with the different parameters (I was much comfortable after studying the original notebook), I managed to achieve a ~0.83 score. After reading a comment from @cdeotte 's (couldn't find the link), I played with the input parameters again (like n_mel, hop_length) and increased my score to ~0.85. I was using effnet_b0 for all of my experiments until then. I switched to b2, b3, b4 and had a score improvement of 0.863, 0.873, 0.88 respectively.

\n

Then, I did rank-based bagging of different models produced along the way and achieved a 0.90 score in the public LB. Finally, inspired by this discussion I used rankdata from the scipy for the ranking part and improved my score to 0.902 for my final submission in the public LB.

\n

Here's also my parameters:

\n

For the SED models:

\n\n

For the standart image models:

\n\n

I used only TP data for training. Tried to add FP data as well but couldn't manage it to work, unfortunately.

\n

Thanks everyone again for this competition journey. I learned a lot along the way.

\n

Edit:

\n

My models private LB was 0.90990.

\n

After applying @cdeotte 's clever postprocessing method it jumps to 0.94729

\n", "rawMarkdown": "Hi everyone,\n\nFirst of all, congrats to all the winners, and thanks to the host for this interesting challenge.\n\nI'm very happy that I ended up in the same position in the private LB as the public and won a bronze medal 😊😊😊. This was kinda the first competition that I heavily spend time on. So, it means a lot to me 😊\n\nI want to thanks @barnwellguy @hidehisaarai1213 @gopidurgaprasad @cdeotte for their insightful comments and for sharing great ideas.\n\nHere is the summary/story of my solution for the problem:\n\nThis was my first audio competition so I had to do some research before doing any experimentation. I started watching some videos from [this](https://www.youtube.com/playlist?list=PL-wATfeyAMNqIee7cH3q1bh4QJFAaeNv0) video series. After grasping some intuition about the problem, I did some data visualization and tried to understand how the sounds change with different classes.\n\nFor the modeling process, I used [this notebook](https://www.kaggle.com/fffrrt/all-in-one-rfcx-baseline-for-beginners) as my baseline and produced my initial models. I had a ~0.75 score from the initial experimentations. Then I started looking at old audio competitions and wanted to see what other people did. So, I did some research on [Cornell](https://www.kaggle.com/c/birdsong-recognition/overview) competition. In that competition, I saw the @hidehisaarai1213 's solution. The code was available on GitHub and it was very well modularized and documented. I mostly focused on the feature representation part from his code. By using the pcen additional to melspectrogram for the feature representation, I had a ~0.81 score. Then, I wanted to understand how the data augmentation is implemented for the audio data so I opened [this](https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/209206) discussion in the forums. With the @barnwellguy 's help out there, I managed to add the data augmentation to my pipeline and achieved a ~0.84 score. That was the first part of my solution. I was using resnest50 in these experimentations.\n\nThen, I wanted to learn about different modeling strategies and researched a bit more on that part. In the Cornell competition, there were many solutions which were including a SED model. And, there is a great introductory [notebook](https://www.kaggle.com/hidehisaarai1213/introduction-to-sound-event-detection) which was written by @hidehisaarai1213 again πŸ™ I carefully studied this notebook, I'm not going to lie, it was kinda overwhelming at first glance, but I believe that it was worth it. There was also an introductory [notebook](https://www.kaggle.com/gopidurgaprasad/rfcx-sed-model-stater) which was based on @hidehisaarai1213 's notebook written by @gopidurgaprasad in this competition. By playing with the different parameters (I was much comfortable after studying the original notebook), I managed to achieve a ~0.83 score. After reading a comment from @cdeotte 's (couldn't find the link), I played with the input parameters again (like n_mel, hop_length) and increased my score to ~0.85. I was using effnet_b0 for all of my experiments until then. I switched to b2, b3, b4 and had a score improvement of 0.863, 0.873, 0.88 respectively. \n\nThen, I did rank-based [bagging](https://www.kaggle.com/kneroma/rfcx-bagging) of different models produced along the way and achieved a 0.90 score in the public LB. Finally, inspired by this [discussion](https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/214676) I used [rankdata](https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.rankdata.html) from the scipy for the ranking part and improved my score to 0.902 for my final submission in the public LB.\n\nHere's also my parameters:\n\n**For the SED models:**\n* PANN's Loss\n* Optimizer: Adam Optimizer(lr=0.001)\n* Scheduler: Warmup+cosine, cosine_hard_restart\n* SR: 32000\n* n_mels: 196\n* hop_length: 320\n\n**For the standart image models:**\n* BCE Loss\n* Optimizer: Adam Optimizer(lr=0.001)\n* Scheduler: Warmup+cosine, cosine_hard_restart\n* SR: 48000\n* n_mels: 128\n* hop_length: 512\n\nI used only TP data for training. Tried to add FP data as well but couldn't manage it to work, unfortunately. \n\nThanks everyone again for this competition journey. I learned a lot along the way. \n\n**Edit:**\n\nMy models private LB was 0.90990.\n\nAfter applying @cdeotte 's clever postprocessing [method](https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/220389) it jumps to 0.94729 \n\n- MODE 1: 0.92908\n- **MODE 2: 0.94729**\n- MODE 3: 0.94456", "votes": null }, { "id": "1207663", "postDate": "02/18/2021 00:49:44", "content": "

Congrats and good luck on the second one!

", "rawMarkdown": "Congrats and good luck on the second one!", "votes": null }, { "id": "1207678", "postDate": "02/18/2021 00:59:07", "content": "

Thanks a lot 😊

", "rawMarkdown": "Thanks a lot 😊", "votes": null }, { "id": "1207740", "postDate": "02/18/2021 02:30:19", "content": "

Congratulations!

", "rawMarkdown": "Congratulations!", "votes": null }, { "id": "1207780", "postDate": "02/18/2021 03:22:54", "content": "

@snnclsr , Congratulations! It's very nice of you to say that. It was my pleasure. I'm very glad that I was of a little help to your first medal :)

", "rawMarkdown": "snnclsr , Congratulations! It's very nice of you to say that. It was my pleasure. I'm very glad that I was of a little help to your first medal :)", "votes": null }, { "id": "1208415", "postDate": "02/18/2021 09:37:17", "content": "

Congrats friend. Your hard work and achievement inspires!

", "rawMarkdown": "Congrats friend. Your hard work and achievement inspires!", "votes": null }, { "id": "1208950", "postDate": "02/18/2021 15:29:00", "content": "

Well done! Happy Kaggling!

", "rawMarkdown": "Well done! Happy Kaggling!", "votes": null }, { "id": "1209348", "postDate": "02/18/2021 21:01:46", "content": "

Thanks a lot 😊 Updated my topic with the solution

", "rawMarkdown": "Thanks a lot 😊 Updated my topic with the solution", "votes": null }, { "id": "1209350", "postDate": "02/18/2021 21:02:39", "content": "

Thanks a lot 😊 It's really great to hear that. Also updated my topic with the solution

", "rawMarkdown": "Thanks a lot 😊 It's really great to hear that. Also updated my topic with the solution", "votes": null }, { "id": "1209352", "postDate": "02/18/2021 21:04:23", "content": "

You were indeed, thanks a lot 😊 You are awesome. Also congrats to you and your team on your final score. Looking forward to read your solution ✌️

", "rawMarkdown": "You were indeed, thanks a lot 😊 You are awesome. Also congrats to you and your team on your final score. Looking forward to read your solution ✌️", "votes": null }, { "id": "1209360", "postDate": "02/18/2021 21:08:49", "content": "

Thanks a lot 😊 Congratulations on your gold medal and becoming a Master as well. It's a great achievement. I will read your solution in detail as soon as possible. Also updated the post with my solution as well. Thanks again ✌️

", "rawMarkdown": "Thanks a lot 😊 Congratulations on your gold medal and becoming a Master as well. It's a great achievement. I will read your solution in detail as soon as possible. Also updated the post with my solution as well. Thanks again ✌️", "votes": null }, { "id": "1211618", "postDate": "02/20/2021 11:19:55", "content": "

Great work !

", "rawMarkdown": "Great work !", "votes": null }, { "id": "1211620", "postDate": "02/20/2021 11:20:59", "content": "

Thanks a lot my friend :)

", "rawMarkdown": "Thanks a lot my friend :)", "votes": null }, { "id": "1213331", "postDate": "02/22/2021 04:10:27", "content": "

This is also my first competition medal. Thank for your sharing resources, they will be usefull for me in the future. Great work!

", "rawMarkdown": "This is also my first competition medal. Thank for your sharing resources, they will be usefull for me in the future. Great work!", "votes": null }, { "id": "1214266", "postDate": "02/22/2021 17:35:40", "content": "

Thanks a lot, congrats on your achievement as well!

", "rawMarkdown": "Thanks a lot, congrats on your achievement as well!", "votes": null }, { "id": "1219157", "postDate": "02/26/2021 13:42:49", "content": "

great work .can I get a medal after the competition ends? and I want adviceπŸ’œπŸ’š

", "rawMarkdown": "great work .can I get a medal after the competition ends? and I want adviceπŸ’œπŸ’š", "votes": null }, { "id": "1219214", "postDate": "02/26/2021 14:58:13", "content": "

You cannot get a medal or competition ranking after a competition has ended. However, you can still make submissions after the deadline if you want to test your methods, as you can see the private/public scores.

", "rawMarkdown": "You cannot get a medal or competition ranking after a competition has ended. However, you can still make submissions after the deadline if you want to test your methods, as you can see the private/public scores.", "votes": null }, { "id": "1219477", "postDate": "02/26/2021 21:22:30", "content": "

Hi @maozragab

\n

@bigironsphere answered your question very well. Thanks!

\n

The links that I provided are good starters to understand the audio data and the architectures related to them. Highly recommended

", "rawMarkdown": "Hi @maozragab \n\n@bigironsphere answered your question very well. Thanks!\n\nThe links that I provided are good starters to understand the audio data and the architectures related to them. Highly recommended", "votes": null } ], "comments": [ { "id": 1207663, "author_name": "flrotm", "author_url": "", "post_date": "02/18/2021 00:49:44", "content": "

Congrats and good luck on the second one!

", "votes": null, "replies": [ { "id": 1207678, "author_name": "snnclsr", "author_url": "", "post_date": "02/18/2021 00:59:07", "content": "

Thanks a lot 😊

", "votes": null, "replies": [] } ] }, { "id": 1207740, "author_name": "bigironsphere", "author_url": "", "post_date": "02/18/2021 02:30:19", "content": "

Congratulations!

", "votes": null, "replies": [ { "id": 1209360, "author_name": "snnclsr", "author_url": "", "post_date": "02/18/2021 21:08:49", "content": "

Thanks a lot 😊 Congratulations on your gold medal and becoming a Master as well. It's a great achievement. I will read your solution in detail as soon as possible. Also updated the post with my solution as well. Thanks again ✌️

", "votes": null, "replies": [] } ] }, { "id": 1207780, "author_name": "barnwellguy", "author_url": "", "post_date": "02/18/2021 03:22:54", "content": "

@snnclsr , Congratulations! It's very nice of you to say that. It was my pleasure. I'm very glad that I was of a little help to your first medal :)

", "votes": null, "replies": [ { "id": 1209352, "author_name": "snnclsr", "author_url": "", "post_date": "02/18/2021 21:04:23", "content": "

You were indeed, thanks a lot 😊 You are awesome. Also congrats to you and your team on your final score. Looking forward to read your solution ✌️

", "votes": null, "replies": [] } ] }, { "id": 1208415, "author_name": "rajkumarl", "author_url": "", "post_date": "02/18/2021 09:37:17", "content": "

Congrats friend. Your hard work and achievement inspires!

", "votes": null, "replies": [ { "id": 1209350, "author_name": "snnclsr", "author_url": "", "post_date": "02/18/2021 21:02:39", "content": "

Thanks a lot 😊 It's really great to hear that. Also updated my topic with the solution

", "votes": null, "replies": [] } ] }, { "id": 1208950, "author_name": "muhakabartay", "author_url": "", "post_date": "02/18/2021 15:29:00", "content": "

Well done! Happy Kaggling!

", "votes": null, "replies": [ { "id": 1209348, "author_name": "snnclsr", "author_url": "", "post_date": "02/18/2021 21:01:46", "content": "

Thanks a lot 😊 Updated my topic with the solution

", "votes": null, "replies": [] } ] }, { "id": 1211618, "author_name": "emirkocak", "author_url": "", "post_date": "02/20/2021 11:19:55", "content": "

Great work !

", "votes": null, "replies": [ { "id": 1211620, "author_name": "snnclsr", "author_url": "", "post_date": "02/20/2021 11:20:59", "content": "

Thanks a lot my friend :)

", "votes": null, "replies": [] } ] }, { "id": 1213331, "author_name": "namdvt", "author_url": "", "post_date": "02/22/2021 04:10:27", "content": "

This is also my first competition medal. Thank for your sharing resources, they will be usefull for me in the future. Great work!

", "votes": null, "replies": [ { "id": 1214266, "author_name": "snnclsr", "author_url": "", "post_date": "02/22/2021 17:35:40", "content": "

Thanks a lot, congrats on your achievement as well!

", "votes": null, "replies": [] } ] }, { "id": 1219157, "author_name": "maozragab", "author_url": "", "post_date": "02/26/2021 13:42:49", "content": "

great work .can I get a medal after the competition ends? and I want adviceπŸ’œπŸ’š

", "votes": null, "replies": [ { "id": 1219214, "author_name": "bigironsphere", "author_url": "", "post_date": "02/26/2021 14:58:13", "content": "

You cannot get a medal or competition ranking after a competition has ended. However, you can still make submissions after the deadline if you want to test your methods, as you can see the private/public scores.

", "votes": null, "replies": [] }, { "id": 1219477, "author_name": "snnclsr", "author_url": "", "post_date": "02/26/2021 21:22:30", "content": "

Hi @maozragab

\n

@bigironsphere answered your question very well. Thanks!

\n

The links that I provided are good starters to understand the audio data and the architectures related to them. Highly recommended

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1207659": "Hi everyone,\n\nFirst of all, congrats to all the winners, and thanks to the host for this interesting challenge.\n\nI'm very happy that I ended up in the same position in the private LB as the public and won a bronze medal 😊😊😊. This was kinda the first competition that I heavily spend time on. So, it means a lot to me 😊\n\nI want to thanks @barnwellguy @hidehisaarai1213 @gopidurgaprasad @cdeotte for their insightful comments and for sharing great ideas.\n\nHere is the summary/story of my solution for the problem:\n\nThis was my first audio competition so I had to do some research before doing any experimentation. I started watching some videos from [this](https://www.youtube.com/playlist?list=PL-wATfeyAMNqIee7cH3q1bh4QJFAaeNv0) video series. After grasping some intuition about the problem, I did some data visualization and tried to understand how the sounds change with different classes.\n\nFor the modeling process, I used [this notebook](https://www.kaggle.com/fffrrt/all-in-one-rfcx-baseline-for-beginners) as my baseline and produced my initial models. I had a ~0.75 score from the initial experimentations. Then I started looking at old audio competitions and wanted to see what other people did. So, I did some research on [Cornell](https://www.kaggle.com/c/birdsong-recognition/overview) competition. In that competition, I saw the @hidehisaarai1213 's solution. The code was available on GitHub and it was very well modularized and documented. I mostly focused on the feature representation part from his code. By using the pcen additional to melspectrogram for the feature representation, I had a ~0.81 score. Then, I wanted to understand how the data augmentation is implemented for the audio data so I opened [this](https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/209206) discussion in the forums. With the @barnwellguy 's help out there, I managed to add the data augmentation to my pipeline and achieved a ~0.84 score. That was the first part of my solution. I was using resnest50 in these experimentations.\n\nThen, I wanted to learn about different modeling strategies and researched a bit more on that part. In the Cornell competition, there were many solutions which were including a SED model. And, there is a great introductory [notebook](https://www.kaggle.com/hidehisaarai1213/introduction-to-sound-event-detection) which was written by @hidehisaarai1213 again πŸ™ I carefully studied this notebook, I'm not going to lie, it was kinda overwhelming at first glance, but I believe that it was worth it. There was also an introductory [notebook](https://www.kaggle.com/gopidurgaprasad/rfcx-sed-model-stater) which was based on @hidehisaarai1213 's notebook written by @gopidurgaprasad in this competition. By playing with the different parameters (I was much comfortable after studying the original notebook), I managed to achieve a ~0.83 score. After reading a comment from @cdeotte 's (couldn't find the link), I played with the input parameters again (like n_mel, hop_length) and increased my score to ~0.85. I was using effnet_b0 for all of my experiments until then. I switched to b2, b3, b4 and had a score improvement of 0.863, 0.873, 0.88 respectively. \n\nThen, I did rank-based [bagging](https://www.kaggle.com/kneroma/rfcx-bagging) of different models produced along the way and achieved a 0.90 score in the public LB. Finally, inspired by this [discussion](https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/214676) I used [rankdata](https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.rankdata.html) from the scipy for the ranking part and improved my score to 0.902 for my final submission in the public LB.\n\nHere's also my parameters:\n\n**For the SED models:**\n* PANN's Loss\n* Optimizer: Adam Optimizer(lr=0.001)\n* Scheduler: Warmup+cosine, cosine_hard_restart\n* SR: 32000\n* n_mels: 196\n* hop_length: 320\n\n**For the standart image models:**\n* BCE Loss\n* Optimizer: Adam Optimizer(lr=0.001)\n* Scheduler: Warmup+cosine, cosine_hard_restart\n* SR: 48000\n* n_mels: 128\n* hop_length: 512\n\nI used only TP data for training. Tried to add FP data as well but couldn't manage it to work, unfortunately. \n\nThanks everyone again for this competition journey. I learned a lot along the way. \n\n**Edit:**\n\nMy models private LB was 0.90990.\n\nAfter applying @cdeotte 's clever postprocessing [method](https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/220389) it jumps to 0.94729 \n\n- MODE 1: 0.92908\n- **MODE 2: 0.94729**\n- MODE 3: 0.94456", "1207663": "Congrats and good luck on the second one!", "1207678": "Thanks a lot 😊", "1207740": "Congratulations!", "1207780": "snnclsr , Congratulations! It's very nice of you to say that. It was my pleasure. I'm very glad that I was of a little help to your first medal :)", "1208415": "Congrats friend. Your hard work and achievement inspires!", "1208950": "Well done! Happy Kaggling!", "1209348": "Thanks a lot 😊 Updated my topic with the solution", "1209350": "Thanks a lot 😊 It's really great to hear that. Also updated my topic with the solution", "1209352": "You were indeed, thanks a lot 😊 You are awesome. Also congrats to you and your team on your final score. Looking forward to read your solution ✌️", "1209360": "Thanks a lot 😊 Congratulations on your gold medal and becoming a Master as well. It's a great achievement. I will read your solution in detail as soon as possible. Also updated the post with my solution as well. Thanks again ✌️", "1211618": "Great work !", "1211620": "Thanks a lot my friend :)", "1213331": "This is also my first competition medal. Thank for your sharing resources, they will be usefull for me in the future. Great work!", "1214266": "Thanks a lot, congrats on your achievement as well!", "1219157": "great work .can I get a medal after the competition ends? and I want adviceπŸ’œπŸ’š", "1219214": "You cannot get a medal or competition ranking after a competition has ended. However, you can still make submissions after the deadline if you want to test your methods, as you can see the private/public scores.", "1219477": "Hi @maozragab \n\n@bigironsphere answered your question very well. Thanks!\n\nThe links that I provided are good starters to understand the audio data and the architectures related to them. Highly recommended" }, "source": "meta" }