Faster Whisper is a reimplementation of OpenAI’s Whisper speech-recognition model built on CTranslate2 — delivering the same accuracy up to four times faster, with far lower memory use.
Faster Whisper is an open-source project by SYSTRAN, hosted on GitHub with around 26,000 stars and active maintenance in 2026. It runs Whisper models on the CTranslate2 inference engine, which is heavily optimised for Transformer models on CPU and GPU. In the project’s benchmarks, large-v2 transcription is up to four times faster than the original implementation at the same accuracy, and 8-bit quantisation cuts memory further — making it practical to transcribe long audio on modest hardware. The Python API supports batched inference, word-level timestamps, and voice activity detection to skip silence.
The library is deliberately a drop-in engine rather than an application: developers load a Whisper checkpoint — including community conversions and Distil-Whisper — choose a compute type such as float16 or int8, and transcribe files or streams with timestamps. This focus is why faster-whisper quietly powers a large share of the subtitling tools, media pipelines, and meeting transcribers built on top of Whisper today.
Faster Whisper is completely free and open source under the MIT license. You download the model weights and run everything locally — no API keys, usage fees, or cloud required.
Developers and teams building transcription into products, pipelines, or research workflows — podcast and video tooling, meeting notes, subtitle generation — who need Whisper-quality accuracy at production speed and cost.