faster-whisper
A Python reimplementation of OpenAI's Whisper speech-to-text model built on the CTranslate2 inference engine.
At a Glance
faster-whisper is a reimplementation of OpenAI's Whisper using CTranslate2, released under the MIT License and free to use, modify and distribute.
Engagement
Available On
Listed Nov 2026
About faster-whisper
faster-whisper is an open-source Python library from SYSTRAN that reimplements OpenAI's Whisper speech recognition model on top of CTranslate2, a fast inference engine for Transformer models. The README states it is up to 4 times faster than openai/whisper at the same accuracy while using less memory, and that 8-bit quantization can improve efficiency further on CPU and GPU. It is released under the MIT License.
What It Is
faster-whisper is a speech-to-text transcription library. Developers load a Whisper model by size (for example large-v3), run it on CPU or NVIDIA GPU, and pass an audio file to receive transcription segments with start and end times plus the detected language and its probability. Audio is decoded with PyAV, which bundles FFmpeg libraries, so a separate FFmpeg installation is not required.
Transcription Options
The library supports FP16, INT8 and INT8-FP16 compute types. It also offers batched transcription through a BatchedInferencePipeline, word-level timestamps, and a Silero VAD filter that removes non-speech audio. The VAD filter is configurable and is enabled by default for batched transcription. Distil-Whisper checkpoints such as distil-large-v3 are compatible. Segments are returned as a generator, so transcription only runs as they are iterated.
Model Conversion and Loading
Models load automatically as converted CTranslate2 weights from the Hugging Face Hub. A ct2-transformers-converter command can convert original OpenAI or fine-tuned Whisper models compatible with Transformers, with optional quantization. Converted models can then be loaded from a local directory or from the Hub.
Setup and Ecosystem
Installation is via pip and requires Python 3.9 or greater. GPU execution needs cuBLAS for CUDA 12 and cuDNN 9, with documented workarounds for older CUDA versions. The README lists community projects built on it, including OpenAI-compatible servers, command-line clients, diarization tools and real-time transcription tools.
Community Discussions
Be the first to start a conversation about faster-whisper
Share your experience with faster-whisper, ask questions, or help others learn from your insights.
Pricing
Open Source (MIT)
faster-whisper is a reimplementation of OpenAI's Whisper using CTranslate2, released under the MIT License and free to use, modify and distribute.
- MIT License
- Up to 4 times faster than openai/whisper for the same accuracy
- 8-bit quantization on CPU and GPU
- Self-hosted Python library (Python 3.9 or greater)
- FFmpeg does not need to be installed
Capabilities
Key Features
- Whisper reimplementation on CTranslate2
- CPU and NVIDIA GPU inference
- 8-bit quantization (INT8, INT8-FP16) and FP16
- Batched transcription pipeline
- Word-level timestamps
- Silero VAD filter for skipping silence
- Distil-Whisper checkpoint compatibility
- Language detection with probability
- Whisper model conversion via ct2-transformers-converter
- No separate FFmpeg installation required (uses PyAV)
