Here's a raw transcript from the demo above:
"Um, so, take my notes from the, the Q2 pipeline review and uh... turn them into like a one-pager, a branded PDF, for the board meeting tomorrow."
And here's what General Input put in the prompt box:
"Turn the Q2 pipeline review notes into a branded one-page PDF for tomorrow's board meeting."
Voice input has been one of the biggest feature requests since we launched: voice to workflows, voice to apps, voice to queries. It's live today. Tap the mic in the prompt box, talk, and the prompt writes itself. Here's why that's worth caring about, and how the cleanup actually works.
Where the speed comes from
The average person types around 40 words a minute. Conversational speech runs 120 to 150. That gap exists because typing makes you do two jobs at once: compose the thought, then serialize it through your fingers, one keystroke at a time, backspacing over typos as you go. Speech skips the second job entirely. Your thoughts already come out as spoken language; there's no extra motor step between having the sentence and producing it.
This isn't a new observation. In 2016, researchers at Stanford, the University of Washington, and Baidu ran the head-to-head: speech input was 3.0x faster than the keyboard for English, with a 20.4% lower error rate. The researchers expected a slight edge and got a blowout.
Prompts are the worst case for typing. An agent does its best work when you give it context, constraints, and what the output should look like. That's a paragraph or three, many times a day. It's exactly the kind of text where a 3x difference compounds, which is how a dictation app like Wispr Flow ends up reportedly raising at a $2 billion valuation.
Why dictation never stuck
Read the raw transcript at the top again. It has filler, a repeated word, a trailing "uh", and a correction mid-sentence. That's how everyone talks. Older dictation tools transcribed it faithfully, so you either sent embarrassing text or you dictated like a court stenographer, speaking punctuation out loud in clean, pre-planned sentences. Planning your sentences before saying them burns the entire speed advantage. Transcription accuracy got solved years ago; the ramble problem is what kept keyboards in charge.
What happens after you stop talking
Two systems run back to back.
While you talk, the browser streams your audio directly to AssemblyAI over a short-lived token our server mints for the session. The audio never passes through General Input's servers. Words appear as you say them.
When you stop, the finished transcript and whatever was already in the prompt box go to Gemini with one job: produce the text you meant. Concretely, it:
- Drops filler and false starts, and resolves self-corrections down to the corrected version. "The, the Q2 pipeline review" becomes "the Q2 pipeline review"; "make it for Thursday, actually Friday" becomes "Friday."
- Fixes punctuation, casing, and obvious mis-transcriptions.
- Applies spoken instructions instead of transcribing them. Say "delete that last sentence" or "make it shorter" and it edits the draft rather than writing the words "delete that last sentence" into your prompt.
- Weaves the new speech into whatever you'd already typed, and answers in the language you spoke.
That third behavior is the part we like most. It means the mic isn't just dictation, it's editing. You can draft a prompt by voice, look at it, and keep talking to revise it.
And if the cleanup model is ever unavailable, the raw transcript still lands in the box. You lose the polish, never the words.
Closer to the speed of thought
Building on General Input was already 10x faster than assembling the same automation in Zapier, n8n, or Make. The keyboard was the next slowest link. Prompts are the highest-leverage text you write all day. Stop typing them.