This code implements finetuning of OpenAI Whisper models using Pytorch Lightning. Most of the code is inspired by (and partly directly copied from) whisper-finetuning. However, since the current code is based on Pytorch Lightning, it also support multi-GPU training. It also supports training with SpecAugment (based on the implmenetation in ESPNet).

The finetuning method implemented here (and also the one in whisper-finetuning) is quite different from the finetuning code in HuggingFace Transformers and ESPNet, and is more similar to the training method that was actually used for training Whisper (according to the paper): finetuning is done on 30-second chunks extracted from long audio recordings, with intra-utterance timestamps. Therefore, the resulting model works very well for transcribing long audios, using e.g. faster-whisper.

TODO