Transcribe

Listens through the microphone and transcribes what was said — with an AI model, which can also answer a question about it, or on the phone itself for free.

An action in the AI group. Its type is action.transcribe.

Declared ports

Port	Label	Flow	Type
in	in	execution in	—
out	out	execution out	—
prompt	What to ask	data in	String
answer	Answer	data out	String
language	Language	data out	String

Configuration

Field	Key	Default	Notes
Transcribe with	using	ai	ai, phone
Model	modelRef	—	chosen from AI_MODEL; shown when using is ai
What to ask	prompt	—	leave empty to transcribe it; shown when using is ai
Language	languageMode	phone	phone, detect, chosen; shown when using is phone
Which language	language	—	chosen from RECOGNITION_LANGUAGE; shown when languageMode is chosen
Listen for at most	maxSeconds	15	seconds
Stop after this much silence	silenceSeconds	3	seconds, 0 = listen the whole time
Longest reply	maxOutputTokens	1024	tokens; shown when using is ai
If it fails	fallback	—	—

Needs

• microphone access

A missing permission is flagged in the editor rather than blocking the macro — the graph is fine, the phone is not, and granting it starts the node working with no edit here at all.

Notes

Transcribe opens the microphone, sends what it heard to an AI model, and puts the model’s answer on the answer port. Nothing is written to storage at any point.

• Listen for at most (seconds) is required and is capped at 120 by the app.
• Stop after this much silence ends the clip early once somebody has spoken. Zero listens for the whole time.
• What to ask empty means transcribe. Filled in, it asks about what was said.

Working of the Node

• The node claims the microphone. If a Start Recording is already running it reports that and stops there, because there is one microphone.
• It records mono 16 kHz WAV into memory, watching how loud each moment is.
• Once it has heard speech, a stretch of quiet as long as Stop after this much silence ends the clip. Quiet before the first word does not count, so a slow speaker is not cut off.
• The clip goes to the model, and the reply lands on answer.
• The clip is discarded either way. There is no file to clean up.

Example: a spoken yes-or-no gate

• Add Transcribe. Set Listen for at most to 8 and Stop after this much silence to 2.
• Set What to ask to Answer only YES or NO: did the speaker agree?
• Wire answer into an If node comparing against YES.
• Wire the If node’s true branch into whatever the agreement should do.

Explanation:

• The question turns a recording into a decision, which a plain transcript would still need parsing to reach.
• Two seconds of silence ends the clip as soon as the speaker stops, so the macro does not wait out the full eight.

Choosing an Engine

Transcribe with decides who does the work, and the two are not interchangeable.

• An AI model sends the sound to the model you picked. It costs a request, needs a network, understands any language the model does, and is the only option that can answer a question about what was said.
• This phone, offline uses the phone’s own recogniser. It is free, needs no key and no signal, and only ever gives the words back — so the question and reply-limit fields are hidden when it is chosen.
• Setting Stop after this much silence to 0 keeps listening through pauses for the whole time, on the phone engine as well as the AI one.
• Offline needs the language pack for the language spoken to be installed. The form says so, and a run that finds it missing names it in the console.

Choosing a Language

Language has three answers, and only one of them needs you to know a language code.

• The phone’s language is the default and matches whatever the phone is set to.
• Detect automatically lets the recogniser work out what is being spoken and switch to it. It needs Android 14 and the packs for the languages involved; older phones fall back to the phone’s language and say nothing.
• A language I choose reveals a chooser listing the packs installed on this phone, each under its own name.

The language port carries the code the recogniser says it used, so a macro can branch on it — reply in whichever language somebody spoke, say. It is empty when nothing reported one, which is every AI transcription and every phone older than Android 14.

How It Differs From Listen

Both nodes hear one spoken answer. They use different ears, and the choice is real.

• Listen uses the phone’s own recogniser. It is free, it can work with no network, and it only knows the languages this phone shipped with.
• Transcribe sends the sound to a model. It costs a request, it needs a network, and it understands any language the model does.
• Only this node can answer a question. Listen returns words; this returns whatever you asked for.
• Listen costs nothing, and so does this node set to This phone, offline — they use the same recogniser. Listen routes on three exec ports where this one puts the fallback on its single port, which is the only difference left.
• Listen routes on three exec ports for heard, timed out and nothing. This node has one out port and puts If it fails on the port when nothing was said.

Why It Records WAV

• Record Audio writes MPEG-4, which suits a file you keep and mail to somebody.
• Gemini refuses that format outright, and OpenAI’s question path takes WAV and MP3 only.
• Mono 16 kHz WAV is accepted by every provider that can hear, so this node works on any connection that works at all.
• The format costs about 32 KB per second, which is why the length is capped.

Points to Remember

• Nothing was said is not a failure. The node writes one line to the console and puts If it fails on the port, so a macro listening on a schedule does not fill the console with errors overnight.
• The microphone grant is the same one the Record Audio nodes use. Granting it once covers all of them.
• The system’s recording indicator is lit for the whole clip. That is Android, not the app.
• Claude cannot hear. Point Model at Gemini, OpenAI, OpenRouter or a self-hosted model.
• Stopping the macro mid-clip releases the microphone. No recording is left half-open.
• This node holds the run for the whole span. Use Start Transcribing and Stop Listening with AI when the macro has something to do while it listens.