Quick pronunciation note before anything else: GOESB = “goose-bee.” Yes, it’s a bit silly. We’re fine with that.
So here’s the actual story of why this exists: every time someone asks “what hardware do I need for local Whisper,” you get a pile of replies like “runs great on my Pi 5!” or “my N100 handles it fine” — and there’s no way to actually check any of it. Different audio, different settings, no idea if the number’s even real. You’re just trusting a stranger’s vibes on the internet, which is a weird way to make a hardware decision. (There’s actually a good example of this right here on this forum — someone posted their own ad-hoc Faster Whisper benchmark script a while back, born from the exact same frustration. Good effort, but a one-off script on one person’s machine isn’t something you can compare against your own setup.)
So I got annoyed enough to build GOESB:
- You run it, but we don’t just take your word for it. Every result gets its hash checked and cryptographically signed by our server before it’s accepted. No “trust me bro” numbers, no cherry-picking your best run.
- Nothing leaves your machine except the numbers. Audio, models, all local. We never see or want your data — we just want to know how fast/accurate it ran and on what.
- No favorites. Works across faster-whisper, vosk, and whisper.cpp, across basically anything from a Pi to a Jetson to a desktop GPU. We’re not trying to sell you one engine.
End goal: an actual, honest leaderboard you can filter by your language/hardware/budget, instead of scrolling through fifteen forum threads hoping someone with your exact setup chimed in.
Takes about 2 minutes to try it yourself:
curl -fsSL https://raw.githubusercontent.com/taktx-io/GOESB/main/scripts/install.sh | sh
goesb
or
pipx install goesb-runner
goesb
Interactive picker walks you through it, runs a real benchmark on your box, offers to submit. If you’re on a Pi, an N100 box, an old laptop, whatever — that data is exactly what’s missing right now.
Heads up on tiny/older boards: bigger models (medium/large-v3) need real RAM, and we don’t yet warn you before a model just doesn’t fit — worst case it gets OOM-killed by the OS mid-run rather than failing cleanly. Stick to tiny/base/small if you’re on something like a Pi Zero or an old low-RAM board.
The runner is open source, and we’d love contributors — found a bug (and since this is a really new project we can expect some), want an engine or platform we don’t support yet, or just want to poke at the code? Everything (feature requests, issues, PRs) goes on the GitHub repo: GitHub - taktx-io/GOESB: Open Edge Speech Benchmark · GitHub
Leaderboard, docs, and everything else lives at: https://www.goesb.com