faster-whisper

SYSTRAN maintains faster-whisper, which runs OpenAI's Whisper speech recognition models on CTranslate2, a C++ Transformer inference engine, in place of PyTorch. The payoff is speed and memory: the project's own benchmark transcribes 13 minutes of audio with large-v2 on an RTX 3070 Ti in about a minute at FP16, and in roughly 16 seconds with batched INT8 inference. You load a model by size name, from tiny through large-v3 and turbo plus distil-whisper checkpoints, and converted weights download from Hugging Face automatically; CTranslate2's converter handles your own fine-tuned Transformers checkpoints. Transcription yields timestamped segments lazily, with language detection, optional word timings, and a Silero voice activity filter that drops silence. PyAV handles decoding, so no system FFmpeg is required. It needs Python 3.9 or newer; CPU works, and GPU runs need NVIDIA's CUDA 12 cuBLAS and cuDNN 9 libraries. The code is MIT, as are the original Whisper weights, so commercial products can embed it. WhisperX, WhisperLive, and the speaches server build on it. It fits developers adding offline transcription to apps or pipelines.