Transcribe File

Sends a sound file to an AI model and returns the transcript, or the answer to a question about what was said.

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

Declared ports

Port	Label	Flow	Type
in	in	execution in	—
out	out	execution out	—
audio	Audio file	data in	String
prompt	What to ask	data in	String
answer	Answer	data out	String

Configuration

Field	Key	Default	Notes
Model	modelRef	—	chosen from AI_MODEL
Audio file	audio	—	—
What to ask	prompt	—	leave empty to transcribe it
Longest reply	maxOutputTokens	1024	tokens
If it fails	fallback	—	—

Notes

Transcribe File reads a sound file from the phone and puts an AI model’s answer about it on the answer port. Leave the question empty and you get a transcript; fill it in and you get an answer about what was said.

• Audio file takes a path, and is wired far more often than typed. The recording worth transcribing is usually the one action.record_audio just made.
• What to ask is the only switch that matters. Empty means transcribe.
• Files are read up to 4 MB. A larger one is refused by name, not truncated.

Working of the Node

• The node reads the file through the same file layer every action.file_* node uses, so a folder you granted months ago still works at three in the morning.
• It works out what the file is from its extension. An extension it does not recognise stops the run there, before anything is sent.
• With What to ask empty, the clip goes to the model provider’s transcription service where one exists, and comes back as text.
• With What to ask filled in, the clip and the question go to the chat model together.
• Whatever comes back lands on answer. Every failure lands on If it fails and the node still pulses out.

Example: mail yourself a transcript of a voice note

• Add a Record Audio node with Seconds set to 30.
• Add Transcribe File and wire Record Audio’s state output into Audio file through a Break Struct on the path field.
• Pick a Model, leave What to ask empty.
• Wire answer into a Send Mail node’s body.

Explanation:

• The empty question is what makes this a transcript rather than a summary.
• The path is wired rather than typed, because Record Audio chooses the file name.

Which Providers Can Hear

Audio support varies more than picture support does, and the node says so before it sends anything rather than after.

• Gemini takes WAV, MP3, AIFF, AAC, OGG and FLAC. It does not take audio/mp4, which is what this app’s own Record Audio node writes. Convert the file, or use Listen with AI, which records WAV.
• OpenAI transcribes almost any format when the question is empty. With a question filled in it takes WAV and MP3 only, and needs a model that accepts audio.
• A self-hosted server transcribes when it serves /audio/transcriptions, which whisper.cpp, faster-whisper and LM Studio all do. Nothing leaves your network.
• OpenRouter needs a model that accepts audio, and always takes the question path.
• Claude cannot hear at all. The node refuses with a sentence saying so.

Points to Remember

• A blank question is a setting, not a missing field. It routes the clip to the service built for transcripts, which is cheaper and available on servers no chat model runs on.
• An unrecognised extension is refused with the file named. Rename it rather than guessing at the model.
• 4 MB is about seventeen minutes of voice recording. The limit protects the phone’s memory, not your quota.
• A transcript longer than Longest reply (tokens) is cut off, reaches the port anyway, and says so in the console.
• The node has no tool switch. Sound reaches a model on the single-question path only.
• This is the one node in the family with no offline option, and it is the platform’s limit rather than a gap. The phone’s recogniser listens to a live microphone and cannot be pointed at a file: asked to read one, it opens the microphone instead and transcribes the room, with no error to say it did. Transcribe and Start Transcribing offer the offline engine because they are listening live.