pyannote-audio VS segmentation_models.pytorch

pyannote/pyannote-audio

Open Source. I've found this to be pretty nice, which is just a wrapper on some hugging face models https://github.com/pyannote/pyannote-audio

(not the creator, but I've built something similar for personal use)

This is a great library for determining which speaker is speaking during each time in an audio file (this is called speaker diarization); I imagine they used it or something like it. Works really well out of the box!

https://github.com/pyannote/pyannote-audio

Do you think pyannote could be implemented in the Pro version of the app to support diarization?

Free startup idea: Use Whisper with pyannote-audio[0]’s speaker diarization. Upload a recording, get back a multi-speaker annotated transcription.

Make a JSON API and I’ll be your first customer.

[0] https://github.com/pyannote/pyannote-audio

Whisper can’t, but pyannote-audio can. I’ve seen a couple of prototypes out there which link the two together.

You can use pyannote python library. It will identify different speakers from audio and will create small audio files with those speakers.

This Colab notebook is basically a standard template (with small changes) provided by pyannote-audio, the library implementing the speaker diarization functionality we need. (template)

Also, I’d suggest considering switching to the segmentation-models library - it provides U-Net models with a variety of pretrained backbones of as encoders. The author also put out a PyTorch version. https://github.com/qubvel/segmentation_models.pytorch https://github.com/qubvel/segmentation_models

SMP offers a wide variety of segmentation models with the option to use pre-trained weights.

SMP has a lot of different choices for architecture other than unet, and a ton of different encoders. I like deeplabv3+/unet with regnety encoder, works well for most things https://github.com/qubvel/segmentation_models.pytorch

This basic example from segmentation models PyTorch repo would be good tutorial to start with. The library is very good, I like the unet, fpn and deeplabv3+ architectures with regnety as encoder https://github.com/qubvel/segmentation_models.pytorch/blob/master/examples/binary_segmentation_intro.ipynb

Sounds like a good semantic segmentation problem, I like this repo: https://github.com/qubvel/segmentation_models.pytorch

When i pass the CT-Scans and the masks to the Loss Function, which is the Jaccard-Loss from the segmentation_models.pytorch library, the value does not decrease but stay in the range of 1.0-0.9 over 50 epochs training on only one batch of 32 images. As far as I have understood, my network should overfit and the loss should decrease since I am only training on one batch of a small amount of images. However this does not happen. I also tried more batches with all the data over 100 epochs, but the loss does not decrease either obviously. Does anyone have an idea what I might have done wrong? Do I have to change anything when passing the masks to my loss function?

I know there is a decent one in segmentation models python (MA-Net: A Multi-Scale Attention Network for Liver and Tumor Segmentation)

You could also use qubvel's segmentation models if you would like to explore semantic segmentation.

Check out segmentation models pytorch, really easy to use, has a great interface.

Have you tried another implementation? For example: qubvel/segmentation_models.pytorch