Faster-Whisper

Faster-Whisper is SYSTRAN's high-performance inference fork of Whisper, first released in February 2023, with ★12k+ on GitHub, licensed under MIT, and written primarily in Python (CTranslate2 backend). Without loss of accuracy, it achieves 4x real-time inference speed on CPU and 4-5x acceleration on GPU for Whisper large-v3 through CTranslate2 quantization and operator optimization, with VRAM consumption reduced by ~50% (FP16 quantization). 100% API-compatible with the original Whisper interface, it's the production-first fork — VisLane's already-listed whisper is the SOTA reference model, faster-whisper is the deployment-first choice.

Core capabilities center on "lossless acceleration + interface compatibility". Backend: CTranslate2 (a C++ library for Transformer inference, supports INT8/INT16/FP16 quantization). Models: supports the full Whisper series (tiny / base / small / medium / large-v1 / large-v2 / large-v3 / distil-large-v3). Speed: large-v3 is 4-5x real-time on RTX 3090, 3-4x real-time on Apple Silicon M1, 1-2x real-time on CPU (8 cores). VRAM: large-v3 INT8 quantized ~3GB VRAM (original ~10GB). Interface: 100% compatible with openai-whisper (from faster_whisper import WhisperModel), directly swappable. Extra features: beam search + word-level timestamps + VAD filter + language detection. Ecosystem: HuggingFace Transformers / faster-distil-whisper / faster-whisper-large-v3-turbo variants; behavior consistent with OpenAI API.

Use cases include: production ASR service (original Whisper inference too slow), batch video subtitle generation, podcast batch transcription, voice search backend, embedded device (CPU/edge) real-time transcription, low-VRAM GPU (consumer 8GB) running large-v3. Hardware floor: CPU is enough (INT8 quantization lets large-v3 run 1x real-time on laptop CPU); GPU optional (8GB+ VRAM runs large-v3 INT8 at 4x real-time). Onboarding: pip install faster-whisper → from faster_whisper import WhisperModel; model = WhisperModel('large-v3', device='cpu', compute_type='int8'); segments, info = model.transcribe('audio.mp3'); CLI compatible with openai-whisper interface (faster-whisper audio.mp3 --model large-v3 --language Chinese --output_format srt).

Licensed under MIT, allows commercial closed-source use, modification, and redistribution. Positioning versus VisLane's already-listed whisper (OpenAI original): Whisper is "academic reference SOTA model" (PyTorch original); Faster-Whisper is "production deployment-first fork" (CTranslate2 backend + quantization); the two are 100% API-compatible, directly swappable. Whisper fits research / teaching / small batch; Faster-Whisper fits SaaS / high-concurrency / edge deployment / VRAM-constrained scenarios. Companion ecosystem: Faster-Whisper has been adopted by Kuaishou / ByteDance / NVIDIA Riva / OpenAI Whisper API backend and other production projects.