Whisper-Finetune

Why we included this project

Whisper-Finetune is the kind of toolkit you reach for when OpenAI's stock Whisper checkpoints don't quite fit your domain. It handles the whole pipeline in one place: preparing training data such as AIShell or WenetSpeech, LoRA fine-tuning on models from tiny up to large-v3-turbo, merging the adapters back in, evaluating character error rate, and converting the result to CTranslate2 or GGML for faster inference. The data handling is its real strength, since you can train with timestamps, without them, or even with transcripts and no audio at all, which matters when your dataset is messy or you only have text to work from. It also goes beyond training, shipping a server API, a Windows desktop app, and an Android demo so a tuned model can actually reach users instead of staying in a notebook.