faster-whisper

Why we included this project

faster-whisper is the reimplementation of OpenAI's Whisper that most production teams end up using. It runs the same model weights on CTranslate2, which the project's benchmarks show to be up to four times faster than the original for the same accuracy, with lower memory use. The 8-bit quantization option makes large models practical on CPU-only hardware. The API mirrors the original Whisper interface closely, so code written against openai/whisper ports over with minimal changes, and audio decoding happens internally through PyAV, so there's no separate FFmpeg install to manage. For meeting transcription, subtitle generation, or call analytics, the batched inference path is where the throughput gains show up.