Has anyone had any luck training with data augmentation strategies? I know data augmentation is really helpful in CV (unless some new research paper I haven't read has overturned that), but I was surprised (maybe due to my naivety in DL audio) to find out audio had so many augmentation types.
\nI slapped together a notebook, basically replicating The Sound of AI's tutorial. For a sample on LibriSpeech, seems the distortions aren't too bad, but there are enough artifacts for me to question efficacy. CTC loss seems to decrease when using them, but I haven't trained on a mass scale.
\nMaybe someone who's a native speaker can attest to the quality of the Bengali audio when transformed?
\nhttps://www.kaggle.com/code/msthil/bengali-data-augmentation
", "messages": [ { "id": "2467447", "postDate": "10/04/2023 14:57:27", "content": "Has anyone had any luck training with data augmentation strategies? I know data augmentation is really helpful in CV (unless some new research paper I haven't read has overturned that), but I was surprised (maybe due to my naivety in DL audio) to find out audio had so many augmentation types.
\nI slapped together a notebook, basically replicating The Sound of AI's tutorial. For a sample on LibriSpeech, seems the distortions aren't too bad, but there are enough artifacts for me to question efficacy. CTC loss seems to decrease when using them, but I haven't trained on a mass scale.
\nMaybe someone who's a native speaker can attest to the quality of the Bengali audio when transformed?
\nhttps://www.kaggle.com/code/msthil/bengali-data-augmentation
", "rawMarkdown": "Has anyone had any luck training with data augmentation strategies? I know data augmentation is really helpful in CV (unless some new research paper I haven't read has overturned that), but I was surprised (maybe due to my naivety in DL audio) to find out audio had so many augmentation types.\n\nI slapped together a notebook, basically replicating [The Sound of AI's](https://www.youtube.com/playlist?list=PL-wATfeyAMNoR4aqS-Fv0GRmS6bx5RtTW) tutorial. For a sample on LibriSpeech, seems the distortions aren't too bad, but there are enough artifacts for me to question efficacy. CTC loss seems to decrease when using them, but I haven't trained on a mass scale.\n\nMaybe someone who's a native speaker can attest to the quality of the Bengali audio when transformed?\n\n[https://www.kaggle.com/code/msthil/bengali-data-augmentation ](url)", "votes": null }, { "id": "2467510", "postDate": "10/04/2023 16:05:06", "content": "Data augmentation can indeed be beneficial for audio data, introducing variations to improve model robustness. However, it's crucial to strike a balance, as excessive augmentation may introduce artifacts. If you've had success or insights, please consider upvoting to share your experiences and foster discussions on effective audio data augmentation strategies. Thank you!
", "rawMarkdown": "Data augmentation can indeed be beneficial for audio data, introducing variations to improve model robustness. However, it's crucial to strike a balance, as excessive augmentation may introduce artifacts. If you've had success or insights, please consider upvoting to share your experiences and foster discussions on effective audio data augmentation strategies. Thank you!", "votes": null }, { "id": "2473015", "postDate": "10/07/2023 19:31:06", "content": "May I ask, did you get any gains with this approach?
", "rawMarkdown": "May I ask, did you get any gains with this approach?", "votes": null }, { "id": "2473282", "postDate": "10/08/2023 06:34:57", "content": "Unfortunately, I haven't had the time to thoroughly try this, however, I did do a quick and dirt training of about 1000 steps using the below augmentation parameters and the loss decreased out from (what I assume was) a local minimum I seemed to have been caught in. That decrease though wasn't too crazy. Hoping to have more time over the next few days to go crazy on a bigger GPU. Please do share though if you manage to see some gains.
\naugment = Compose([\n AddGaussianNoise(min_amplitude=, max_amplitude=, p=),\n TimeStretch(min_rate=, max_rate=, p=),\n PitchShift(min_semitones=-, max_semitones=, p=),\n Shift(p=),\n Trim(top_db=, p=)\n ])\nData augmentation can indeed be beneficial for audio data, introducing variations to improve model robustness. However, it's crucial to strike a balance, as excessive augmentation may introduce artifacts. If you've had success or insights, please consider upvoting to share your experiences and foster discussions on effective audio data augmentation strategies. Thank you!
", "votes": null, "replies": [] }, { "id": 2473015, "author_name": "leonardoboulitreau", "author_url": "", "post_date": "10/07/2023 19:31:06", "content": "May I ask, did you get any gains with this approach?
", "votes": null, "replies": [ { "id": 2473282, "author_name": "msthil", "author_url": "", "post_date": "10/08/2023 06:34:57", "content": "Unfortunately, I haven't had the time to thoroughly try this, however, I did do a quick and dirt training of about 1000 steps using the below augmentation parameters and the loss decreased out from (what I assume was) a local minimum I seemed to have been caught in. That decrease though wasn't too crazy. Hoping to have more time over the next few days to go crazy on a bigger GPU. Please do share though if you manage to see some gains.
\naugment = Compose([\n AddGaussianNoise(min_amplitude=, max_amplitude=, p=),\n TimeStretch(min_rate=, max_rate=, p=),\n PitchShift(min_semitones=-, max_semitones=, p=),\n Shift(p=),\n Trim(top_db=, p=)\n ])\n