Verified project record

SYSTRAN/faster-whisper

faster-whisper delivers up to 4x faster Whisper transcription with lower memory usage using CTranslate2. It supports int8 quantization, voice activity detection, and batched inference on both CPU and CUDA GPUs.

Project overview

It addresses slow inference and high memory consumption in Whisper models by providing a drop-in pipeline with up to 4x faster transcription and 8-bit quantization support.

Project type: Model Runtime · Audio & Speech
Use cases: Meeting Notes · Translation & Subtitles · Audio & Speech
Deployment: Refer to project documentation
License: MIT

Best for

Developers and AI engineers working with speech transcription who need faster inference and lower memory usage on CPU or CUDA GPUs.

Key capabilities

Supports running transcription on both CPU and CUDA GPUs with various compute types including float16 and int8.
Provides a BatchedInferencePipeline drop-in replacement for batched transcription.
Integrates Silero VAD to filter out parts of audio without speech.
Includes a script to convert Transformers-compatible Whisper models to CTranslate2 format.
Generates precise start and end timestamps for individual words during transcription.
Natively supports running Distil-Whisper checkpoints like distil-large-v3.

Limitations and risks

The latest CTranslate2 version only supports CUDA 12 and cuDNN 9.
Requires downloading OpenAI Whisper, CTranslate2, or Distil-Whisper models from Hugging Face Hub.

Getting started

Installable via standard pip command without compiling dependencies locally. The first success path is 'pip install faster-whisper'.

Alternatives and comparisons

Provides real-time audio transcription supporting 100 languages.
Offers speech to text, text to speech, and machine translation using offline engines.
Extracts subtitles from audiobook audio files automatically.

Project comparisons

Evidence and sources

Release: faster-whisper 1.2.1
README: **faster-whisper** is a reimplementation of OpenAI's Whisper model using [CTranslate2](https://github.com/OpenNMT/CTranslate2/), which is a fast inference engine for Transformer m…
README: | faster-whisper | fp16 | 5 | 1m03s | 4525MB | | faster-whisper (`batch_size=8`) | fp16 | 5 | 17s | 6090MB |
README: Unlike openai-whisper, FFmpeg does **not** need to be installed on the system. The audio is decoded with the Python library [PyAV](https://github.com/PyAV-Org/PyAV) which bundles…
README: **Note**: The latest versions of `ctranslate2` only support CUDA 12 and cuDNN 9.

Source updated at: 2026-07-16T18:15:02Z

Open original GitHub project