DAS WhisperSeg

Fork of WhisperSeg for use with DAS.

Install

conda create -n das-wseg python=3.13 uv -y
conda activate das-wseg
uv pip install git+https://github.com/janclemenslab/das-whisper --upgrade

Use

Train/evaluate/predict using the CLI interface:

das-whisper train -h
das-whisper predict -h
das-whisper evaluate -h

To convert your own annotations for use with whisperseg, see the original docs.

Complete usage example:

Train

das-whisper train nccratliri/whisperseg-base-animal-vad models/bf nccratliri/bengalese-finch-subset-with-csv-label/train --val-ratio 0.2 --validate-every 100 --save-every 100

This will download the dataset nccratliri/bengalese-finch-subset-with-csv-label from huggingface, train a whisperseg model using the training set and using the pre-trained model nccratliri/whisperseg-base-animal-vad as a starting point. The best trained model will be saved to models/bf/final_checkpoint.

Evaluate

das-whisper evaluate nccratliri/bengalese-finch-subset-with-csv-label/test models/bf/final_checkpoint

This will evaluate the best model on the test data in the dataset folder.

predict

das-whisper predict nccratliri/bengalese-finch-subset-with-csv-label/test models/bf/final_checkpoint nccratliri/bengalese-finch-subset-with-csv-label/test_predictions.csv

This will use to trained model to create annotations files for the test data in results.