systran/faster-whisper

Faster Whisper is a Python library that converts spoken audio into written text, using a rebuilt version of OpenAI's Whisper speech-recognition model. The key idea is speed: by rebuilding Whisper on top of a faster inference engine called CTranslate2, it can transcribe audio up to four times faster than the original while using less memory. The library works by loading a speech model, pointing it at an audio file, and getting back a stream of timed text segments, essentially time-stamped transcripts. It supports running on a GPU for top speed or on a regular CPU, and it can use a compressed "int8" mode to further cut down memory usage without much accuracy loss. You can also process multiple audio clips at once in a batched mode for even faster throughput. Someone would use this when they need to convert large amounts of audio or video to text quickly, think podcast transcription, meeting notes, subtitle generation, or building a voice assistant. It is also a good fit for anyone who found the original Whisper too slow and wants a drop-in replacement that needs less computing power. The stack is Python, with the CTranslate2 engine under the hood and NVIDIA CUDA for GPU acceleration. Audio decoding is handled internally without needing to install separate tools.