transcribe.cpp

Built on ggml, transcribe.cpp is an inference library that broadly supports the latest speech recognition (ASR) models. It aims to overcome the limitations of existing whisper.cpp or ONNX approaches and provide consistent performance and accuracy across various platforms.

Key Features:

• Extensive model support: Supports 16 ASR families and 60+ models, with continuous expansion.
• Hardware acceleration: Supports acceleration via Vulkan, Metal, CUDA, TinyBLAS, maximizing GPU performance.
• High reliability: Numerical validation and WER (Word Error Rate) testing have been completed for all models, ensuring the same accuracy as reference implementations.
• Diverse functionality: Supports streaming and batch transcription, with high compatibility that allows whisper.cpp's .bin files to be used as-is.
• Developer-friendly: Officially supports bindings for 4 languages: Python, JavaScript/TypeScript, Rust, and ObjC/Swift.