window breakage auto session?

[frankie]

I am using auto session with a segmentation algorithm to detect event in my audio file captured from Quick Feather board. But no matter what segmentation algorithm I used, the accuracy in auto sense result is below 60%. I tried to label the segments manually and use generate auto session to get customized algorithm by training DCL with example events.

But it doesn't work, no idea how to do for the next with limited log message.
I think the root cause is segmentation algorithm has limited maximum segment samples to 8192?!

1-1.jpg (239.64 KiB) Viewed 97333 times

1-2.jpg (119.78 KiB) Viewed 97333 times

1-3.jpg (83.04 KiB) Viewed 97333 times

[Tolido848]

Increase Segment Length: If the segmentation algorithm is limited to 8192 samples, try increasing the segment length if possible. This might help capture more context and improve accuracy.
Data Augmentation: Use data augmentation techniques to artificially increase the size and variability of your training dataset. This can help the algorithm generalize better to new data banana game
Algorithm Tuning: Experiment with different hyperparameters and configurations of your segmentation algorithm. Sometimes, small adjustments can lead to significant improvements in performance

[JoshuaHolland]

I totally understand the frustration with low accuracy rates-audio segmentation can be tricky, especially with limited sample sizes. The 8192 segment limit you mentioned could definitely be a bottleneck. Have you tried experimenting with different audio preprocessing techniques before feeding data to the algorithm? Also, manual labeling for custom training is a solid approach, but make sure your training set is diverse enough. If you're exploring audio-related projects, tools like Heardle can actually help you understand how audio recognition works. Keep iterating with your DCL training-sometimes small adjustments to parameters make a big difference!

[juniap56]

Hey, great post! I've faced similar issues with auto-segmentation accuracy. Have you considered fine-tuning your feature extraction before segmentation? Sometimes, improving the initial representation helps immensely. For DCL training, make sure your manually labeled data is incredibly clean and diverse. Also, what if you tried a different approach for event detection, perhaps something inspired by how people play Uno Online – where each "card" (event) has unique characteristics? Keep us updated on your progress!

[chatterweigh]

Have you thought about optimizing your feature extraction prior to segmentation? Improving the original representation can be very beneficial at times. Make sure your manually labeled data is extremely clean and diverse for DCL training.

[derrickdecker]

It sounds like you’ve already tried several approaches, so I’d start by checking the segmentation settings and the limitations of the algorithm before retraining DCL. The 8,192-sample limit could be relevant if your audio events are longer than the supported segment size, but it would be worth confirming this in the documentation or source code. I’d also check the audio sample rate, event duration, labeling quality, and whether the training examples are representative of the actual events. Since the log messages are limited, comparing a few manually labeled segments with the automatic results might help narrow down where the problem occurs. If anyone has successfully customized an auto session with DCL on the Quick Feather board, sharing the configuration and training steps would be really helpful.