faster-whisper

faster-whisper runs Whisper models through CTranslate2, which can make inference faster and reduce memory use. It keeps the familiar Whisper model sizes, such as tiny, base, small, medium, and large, but adds optimized CPU and GPU execution paths.

Yes, faster-whisper can run on CPU, and quantization options such as int8 help reduce memory and improve speed. A GPU still matters for large models or long audio, but CPU mode is useful for servers and laptops without CUDA.

faster-whisper supports timestamp output, including word-level timestamps in common usage patterns. That is useful for subtitle alignment, search indexing, and transcript editors that need more than segment-level timing.

faster-whisper is an open-source project from SYSTRAN and is commonly used as a practical Whisper inference backend. It is separate from OpenAI's original Whisper repository, even though it runs converted Whisper model weights.