faster-whisper
A reimplementation of OpenAI’s Whisper using CTranslate2, built for faster transcription with less memory. Part of Open-source motion studio: 20 repos.
Checked 2026-10-27 against the GitHub repository, its README, and the npm or PyPI registry where relevant.
| Runtime | Python 3.9+. FFmpeg is not required on the system, because audio is decoded with PyAV, which bundles FFmpeg libraries. GPU use needs cuBLAS and cuDNN 9 for CUDA 12; CPU works. |
| How an agent drives it | Python API; deterministic enough for a captions step (audio in, timed text out) with no GUI. |
| License | MIT. |
| Maintenance | Latest release v1.2.1, 2025-10-31; last commit 2026-10-16; not archived. PyPI faster-whisper 1.2.1. |
What it gives you
- Input: audio files. Output: transcript segments with timestamps;
word_timestamps=Truereturns word-level times (README example). - Batched inference through
BatchedInferencePipeline, 8-bit quantisation on CPU and GPU, and compatibility with distil-whisper checkpoints. - Benchmark quoted from the README (13 minutes of audio, large-v2, GPU, NVIDIA RTX 3070 Ti 8 GB, CUDA 12.4): openai/whisper fp16 2m23s; whisper.cpp fp16 1m05s; faster-whisper fp16 1m03s; faster-whisper batch size 8 fp16 17s. The README claims up to 4x faster than openai/whisper at the same accuracy. These are the maintainers’ figures, not re-measured.
Key differentiator
Same Whisper accuracy with the CTranslate2 engine’s speed and memory savings, plus batched mode and word timestamps. Source: README benchmark section.
Pick something else when
whisper.cpp (MIT) for a C/C++ build with no Python. WhisperX (BSD-2-Clause) when you need stronger word alignment and diarisation. openai-whisper (MIT) as the reference implementation. See Whisper and ASR models.
Agent fit
Strong. A command-line or Python step that produces the timed text the caption blocks of a renderer consume.
Overlap within the 20
Overlaps nothing in the other 19; a natural second stage after Chatterbox or recorded narration, feeding HyperFrames or Remotion captions.