A voice message, step by step
What really happens when you hold the Pilotie microphone and talk: the limits, which speech services write it out, what is kept and what is not.
Voice is the part people ask about first, and the part where honesty matters most. Here is what happens, in order.
1. You hold the button#
The browser records after you have allowed the microphone once. Hold to talk and let go to send, or tap to lock it open when your hands are busy. A message is at most 2 minutes 10 seconds and 6 MB, and a workspace can send 150 a day. A long job is several messages.
2. The clip goes to be written out#
The workspace checks the limits and sends the clip to a speech service. By default that is Groq with Whisper large-v3-turbo. If it fails, OpenAI's gpt-4o-mini-transcribe takes over. The names in your workspace go along as a spelling hint, so a client called Tjerk is not written as Jerk.
3. You see the text#
Your message appears at once with the word Voice and "Transcribing..." until the text is ready. Then the helper answers it like any typed message. If it failed, the message says so.
4. The audio goes away#
The app is set to keep only the text. The audio is not kept after it was read.
What can go wrong#
No microphone, or a blocked one: the page says so and you can type. No speech key on a workspace: the app writes a clearly labelled stand-in text instead of pretending. A noisy van: Whisper copes, but check the text, which is shown to you for a reason.
What is not built yet#
Words appearing while you speak, writing out on the phone itself, and the helper answering out loud are planned, with no date.
A tip for noisy places#
Say names slowly and say the amount twice if it matters. You will see the written-out text before anything is done with it, and a quote or an invoice is only ever a draft until you tap Yes.