
faster-whisper
Faster Whisper transcription with CTranslate2
The Lens
By Erik Loyd, SaaS CEO and former COO/CFO of an AWS Premier Partner.
Updated Nov 2026
faster-whisper runs OpenAI's Whisper speech-to-text models up to four times faster than openai/whisper at the same accuracy, with less memory. It rebuilds inference on CTranslate2 and sits underneath a lot of self-hosted transcription stacks. MIT licensed, nothing gated.
It's a Python library, not a service. Install it, pick a model (tiny through large-v3, plus turbo and distil variants), feed it audio. The README benchmarks it on an 8GB consumer GPU, and int8 makes CPU use practical. GPU setup needs CUDA 12 and cuDNN 9, where most of the pain lives. A server, diarization and live streaming come from community projects like speaches and WhisperX.
Paying AssemblyAI or Deepgram for plain batch transcription? Run the math. AssemblyAI lists $0.15 to $0.21 per audio hour; at thousands of hours, or audio that can't leave your servers, a GPU box wins. Solo: use it. Small teams: worth an afternoon. Larger teams: use it for bulk and private audio, keep an API for live calls.
The catch: maintenance is stop-start. The repo sat nearly silent from late 2025 until October 2026, when fixes and an issue cleanup resumed. But the last release is still v1.2.1 from October 2025, so pip installs a year-old build.
Free vs Self-Hosted vs Paid
fully freeFree: Everything, MIT licensed. Every Whisper model size (tiny through large-v3), turbo and Distil-Whisper models, int8 quantization on CPU and GPU, a batched pipeline, Silero VAD filtering, and word-level timestamps. Nothing is gated.
Self-hosted: pip install faster-whisper; models download on first run. CPU with int8 handles casual use; an NVIDIA GPU with CUDA 12 and cuDNN 9 runs the large models fast (the README benchmarks an 8GB RTX 3070 Ti). The library does transcription only: an HTTP server, speaker diarization and live streaming come from community projects built on it (speaches, WhisperX, whisper_streaming). Expect a day to a working internal endpoint, and you own chunking, retries and scaling.
Paid: None. The commercial comparison is hosted speech-to-text: AssemblyAI lists $0.21/hr for AssemblyAI-STT-3.7 and $0.15/hr for Universal-2 on pre-recorded audio, $0.30/hr for realtime. Deepgram's Nova-3 monolingual streaming is $0.0048/min (about $0.29/hr) on current promo pricing, list $0.0077/min, with pre-recorded cheaper and diarization a $0.0020/min add-on.
The math: At 100 hours a month the APIs cost roughly $15 to $30, not worth self-hosting for. At thousands of hours, or when audio can't leave your infrastructure, a GPU instance wins.
Free MIT-licensed Whisper inference. Beats per-hour APIs on cost at volume or when audio can't leave your servers, but the last release is over a year old.
What to do by team size
- Solo
- free, runs on your own GPU or CPU
- Small team
- free; an afternoon of setup replaces a per-hour API bill
- Medium team
- free for bulk and private audio; keep an API for live streaming
- Large team
- free engine, but pin versions and watch for a long-overdue release
Get tools like this every Wednesday
One featured tool, three on the radar. No fluff.
Similar Tools
A low score is not a verdict on quality. Young and niche tools start low by design. How we calculate scores
Trust Signals
License: MIT License
MIT, no restrictions.
Commercial use: ✓ Yes
About
- Owner
- SYSTRAN (Organization)
- Stars
- 26,934
- Forks
- 2,280
Explore Further
More tools in the directory
rea
Reverse engineer anything with agents, from app behavior down to native binaries.
473.2k ★openclaw
Your own personal AI assistant. Any OS. Any Platform. The lobster way. 🦞
395.0k ★everything-claude-code
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
308.6k ★
