Whisper fine-tuning for ASR

In this Repo, you can easily fine-tune different variations of the Whisper model to your specific multilingual data based on a simple manifest.

1. prepare_data.py :::: to prepare ".csv" files for train and test
2. train.py :::: train and save the fine-tuned Whisper model
3. decode.py :::: decode the test or any evaluation ".wav" file
4. whisper_transcribe_WER.py ::: another (easier) method for utilizing the Whisper model in transcription.

*** You can use different versions of the openai Whisper model.

Auxilary files

The required packages are listed in the "requirements.txt" file and you can easily install all of them using: pip install -r requirements.txt

It would be better to make a new Python environment using python3 -m venv myenv , after that, activate the venv using source myenv/bin/activate and then install the packages.

To run on the servers by Slurm, you can use the slurm_run.sh file.

The "files_test.csv" and "files_train.csv" help us understand better the required files for testing and training.