Instead of OpenAI Whisper: faster-whisper
Use faster-whisper to run Whisper models from Python. It runs the same models up to 4 times faster than the openai-whisper package for the same accuracy, uses less memory, and needs no system FFmpeg. On Apple Silicon, or without Python, use whisper.cpp instead.
Both are for transcribing speech to text with OpenAI's Whisper models.
Why faster-whisper is better
- In faster-whisper's own benchmark on an RTX 3070 Ti, transcribing 13 minutes of audio with the large-v2 model took 2m23s with openai-whisper and 1m03s with faster-whisper, or 17s with batching.
- 8-bit quantization cuts memory. The large-v2 model in int8 used 2926MB of VRAM against 4708MB for openai-whisper in that benchmark.
- On CPU, the small model took 1m42s in int8 with faster-whisper against 6m58s for openai-whisper on the same Intel machine.
- Audio is decoded with PyAV, which bundles the FFmpeg libraries, so there is no system FFmpeg to install.
- Voice activity detection filtering and word-level timestamps are built in as options to transcribe().
When to keep OpenAI Whisper
Keep openai-whisper when you need the reference implementation, for research or to reproduce published results, or when you want to modify or fine-tune the model in PyTorch. It is maintained by OpenAI, though releases are infrequent.
Switching
The code change is usually small. You install faster-whisper, load WhisperModel with a model size and compute type, and call transcribe(), which returns a generator of segments rather than a finished result. Standard model sizes download automatically in CTranslate2 format, and fine-tuned checkpoints need converting with ct2-transformers-converter.
Watch out
faster-whisper depends on CTranslate2, whose current builds support only CUDA 12 and cuDNN 9, so older CUDA setups need pinned older versions. GPU acceleration needs NVIDIA libraries, so Apple Silicon users get CPU speed only. It is maintained by SYSTRAN, and releases come slowly.
Also worth a look
- whisper.cpp A C/C++ port with no Python dependency that runs fully on the GPU on Apple Silicon through Metal; pick it for Macs, desktop apps and embedded use.
- WhisperX Adds word-level timestamps and speaker diarization on top of fast Whisper inference; pick it for subtitles and multi-speaker transcripts.
Sources
Judged on 2026-11-06. See every tools for running AI models locally, or all 66 replacements.