the showdown · real outputs · zero cherry-picking
Other tools transcribe.
FunButton rewrites.
Same messy spoken input. Five scenarios. The left column is what Whisper hears — what most apps paste. The right column is what FunButton sends to your cursor.
Captured live from Groq Whisper Large v3 Turbo + Llama 3.3 70B with the exact prompts in the app source. Re-run yourself: scripts/capture_showdown.sh
The rambling email follow-up
Real meeting context, real fillers, real "um you know".
what we actually said into the mic →
“Hi um Russell I wanted to like you know follow up about the deal we discussed last week and uh I think we should probably move forward but I'm not sure about the pricing maybe we could do something like 30K instead of 25 since uh you know the scope grew a bit. Let me know what you think.”
The Slack message you change mid-sentence
Said one thing, immediately changed your mind. The raw transcript is wrong. The cleanup is right.
what we actually said into the mic →
“Hey can you ship that PR by EOD wait actually no make it tomorrow morning since the build is broken in CI right now”
The arrow function with spoken symbols
"open paren camelCase user data comma options close paren fat arrow…" Try this in Whisper alone.
what we actually said into the mic →
“open paren camelCase user data comma options close paren fat arrow object dot keys open paren user data close paren dot filter open paren camelCase k fat arrow not k dot starts with open paren quote underscore quote close paren close paren”
The function signature with types
Speak the function. Get the syntax. Plus types.
what we actually said into the mic →
“function snake_case fetch user profile takes user id of type number returns promise of user”
The "actually no scratch that" pivot
The wedge in one sample. Nobody else solves this.
what we actually said into the mic →
“Let me draft a quick response to that customer about um the refund situation actually no scratch that I'll handle it tomorrow morning when I have my notes in front of me”
Stop typing. Start talking.
Your voice, your machine, your cursor. Local-first, dev-grade, no API key required (with Ollama).