inference/research
verifiedhigh · 0.90updated 2026-10-04

faster-whisper

cuda, cpu-avx512, cpu-avx2

Fields

id
engines/faster-whisper
Repo
SYSTRAN/faster-whisper
Languages
PythonC++
License
MIT
First release
2023
Design goal
high-performanceCTranslate2-backendbatch-transcriptionease-of-use
Scheduling
none
Backends
cudacpu-avx512cpu-avx2
Formats
safetensors
Notable features
Reimplementation of OpenAI's Whisper using CTranslate2 inference engineUp to 4x faster than openai/whisper for the same accuracy with less memory8-bit quantization (int8) on both CPU and GPU for further speedupBatched transcription pipeline for processing multiple audio files efficientlySupports distil-whisper modelsWord-level timestampsSilero VAD integration for filtering non-speech segmentsOpenAI-compatible server via speaches-ai/speaches
Hardware caveats
GPU execution requires CUDA 12 and cuDNN 9 (or downgrade CTranslate2 for older CUDA)No Metal, Vulkan, or ROCm support -- CUDA or CPU onlyint8 quantization on GPU uses int8_float16 compute type (not pure int8)Batched transcription requires significant VRAM (6GB+ for large-v3 batch_size=8)
Best for
Batch transcription of many audio files on NVIDIA GPUsPython-based workflows that need Whisper accuracy with better speedWord-level timestamps and VAD filtering
Avoid for
Apple Silicon or AMD GPU users (no Metal/Vulkan/ROCm)Production serving with continuous batching (use vLLM or SGLang)Real-time streaming (use whisper-streaming or WhisperLive instead)
Docs
https://github.com/SYSTRAN/faster-whisper
notes
faster-whisper is a Python wrapper around CTranslate2 that reimplements Whisper inference. It is up to 4x faster than the original OpenAI implementation. The project is actively maintained. It does NOT support Metal, Vulkan, or ROCm -- only CUDA and CPU. The batched transcription pipeline is the key feature for throughput. int8 quantization provides additional speedup on both CPU and GPU. ||| MAINTENANCE (dated fact, GitHub API, 2026-10-04): last commit 2026-10-01 (0.1 months before this record), latest release v1.2.1 published 2025-10-31, 47 contributors (distinct committers including anonymous), open_issues_count 61 (issues PLUS pull requests - see [[ev-github-open-issues-count-includes-pull-requests]]), not archived, 25703 stars. See [[ev-maintenance-status-is-four-dated-facts]] for how to read this fact set. ||| DEPLOYMENT: NO EVIDENCE OF PRODUCTION DEPLOYMENT FOUND as of 2026-10-04 - repositories, READMEs, docs and issue trackers were searched. See [[ev-production-evidence-is-scarce-and-nearly-always-single-site]] for the grading scheme; note that for an artifact-evaluation or paper-code release, non-deployment is documented intent rather than a negative finding.

Notes

faster-whisper is a Python wrapper around CTranslate2 that reimplements Whisper inference. It is up to 4x faster than the original OpenAI implementation. The project is actively maintained. It does NOT support Metal, Vulkan, or ROCm -- only CUDA and CPU. The batched transcription pipeline is the key feature for throughput. int8 quantization provides additional speedup on both CPU and GPU. ||| MAINTENANCE (dated fact, GitHub API, 2026-10-04): last commit 2026-10-01 (0.1 months before this record), latest release v1.2.1 published 2025-10-31, 47 contributors (distinct committers including anonymous), open_issues_count 61 (issues PLUS pull requests - see [[ev-github-open-issues-count-includes-pull-requests]]), not archived, 25703 stars. See [[ev-maintenance-status-is-four-dated-facts]] for how to read this fact set. ||| DEPLOYMENT: NO EVIDENCE OF PRODUCTION DEPLOYMENT FOUND as of 2026-10-04 - repositories, READMEs, docs and issue trackers were searched. See [[ev-production-evidence-is-scarce-and-nearly-always-single-site]] for the grading scheme; note that for an artifact-evaluation or paper-code release, non-deployment is documented intent rather than a negative finding.

Cross-references

This record references no other records — it is a leaf. It is cited by 1.