Tools/openai/whisper

whisper

Robust Speech Recognition via Large-Scale Weak Supervision

112.2k ★+452/wkestablishedPythonMIT License →

View on GitHub →Compare in Media Processing →Homepage →

Replaces:AssemblyAI Deepgram Otter.ai Rev

See how it compares →

The Lens

By Erik Loyd, SaaS CEO and former COO/CFO of an AWS Premier Partner.

Updated Oct 2026

Whisper is OpenAI's speech-to-text model, and it's still the engine most open source transcription tools are built on. Feed it audio in any of about 100 languages and you get a transcript, or an English translation. Code and weights are MIT, so running it on your own machine costs nothing.

Setup is pip install -U openai-whisper plus ffmpeg. Hardware is the real cost: tiny and base need about 1GB of VRAM, the default turbo model about 6GB, large about 10GB. Turbo is fast but can't translate. CPU works, slowly. Most people now run the same weights through faster-whisper (up to 4x faster) or whisper.cpp on a Mac.

OpenAI's hosted Whisper API is $0.006 a minute, and its newer gpt-4o-mini-transcribe is half that at $0.003. Solo: hosted for the occasional file, local when the audio is private. Small teams: local on one GPU. Larger teams with real volume: self-host and the per-minute bill disappears.

The catch: this repo gets small fixes but no release since June 2025, it does no speaker labels or live captions, and OpenAI's own model card warns it can hallucinate words nobody said. For a desktop app, use Buzz or Vibe. For speaker labels, WhisperX.

Free vs Self-Hosted vs Paid

fully free

Free: Everything in the repo. The MIT license covers the code and the model weights for all six sizes (tiny, base, small, medium, large, turbo), plus English-only versions of the four smallest. No paid tier, no usage caps.

Self-hosted: pip install -U openai-whisper, plus ffmpeg (and possibly Rust if tiktoken has no prebuilt wheel for your platform). Cost is hardware. Approximate VRAM from the README: tiny and base ~1GB, small ~2GB, medium ~5GB, turbo ~6GB, large ~10GB. The default turbo model is an optimized large-v3 that does not do translation. CPU-only works but is slow beyond the small models. For speed, faster-whisper (MIT) claims up to 4x faster at the same accuracy with less memory, and whisper.cpp (MIT, last release v1.9.5 on 2026-10-06) runs fully on the GPU via Metal on Apple Silicon.

Paid: No paid tier from this project. OpenAI's hosted transcription API, per minute of audio: Whisper $0.006, gpt-4o-transcribe $0.006, gpt-4o-transcribe-diarize (adds speaker labels) $0.006, gpt-transcribe $0.0045, gpt-4o-mini-transcribe $0.003. Live transcription via gpt-realtime-whisper is $0.017/minute. Hosted means your audio leaves your machine; at volume, a GPU box you own wins on cost and privacy.

Free to run locally under MIT; your only cost is a GPU. If you'd rather pay, OpenAI's hosted Whisper is $0.006/minute, but its gpt-4o-mini-transcribe at $0.003/minute is now the cheaper hosted option.

What to do by team size

Solo: free locally; a hosted API at $0.003 to $0.006/minute for the occasional file
Small team: free on one GPU, ideally through faster-whisper
Medium team: free self-hosted; a GPU box beats per-minute billing at volume
Large team: free self-hosted, or a hosted API if you need diarization and don't want to run GPU infra

Self-hosting ops:moderate

View pricing page →

Get tools like this every Wednesday

One featured tool, three on the radar. No fluff.