How yeyupiaoling/whisper-finetune works
Project type and purpose
• This is a combined AI/ML toolkit and full-stack/demo system for fine-tuning OpenAI’s Whisper ASR model, running inference (CLI, server, GUI), and deploying to Android and Windows desktop. It includes:
• Data preparation & augmentation scripts (aishell.py, tools/*)
• Training (finetune.py), merging and evaluation (merge_lora.py, evaluation.py)
• Inference CLI and accelerated variants (infer.py, infer_ct2.py, infer_server.py, infer_gui.py)
• Web frontend (templates/, static/) served by Flask (infer_server.py)
• Mobile app (AndroidDemo/) and desktop app (WhisperDesktop/)
File structure
• Top-level ML scripts: aishell.py, finetune.py, merge_lora.py, evaluation.py, infer*.py
• Configuration: configs/, requirements.txt, run.sh
• Data: dataset/, tools/ (data conversion, speed tests)
• Web UI: templates/ (HTML), static/ (CSS, JS)
• Android client: AndroidDemo/ (Gradle project, JNI wrapper for whisper.cpp)
• Desktop client: WhisperDesktop/ (Windows .exe)
• Helpers: utils/, metrics/
README for additional insights
• It describes training, conversion to CTranslate2 and GGML, evaluation, inference options, GUI and server deployment, Android and Windows demos.
• It lists dependency stack (Python, PyTorch, CTranslate2, faster-whisper, whisper.cpp).