The input you already have
HUUTO lists every CoreAudio device that has input channels. No brand is filtered out, and no hardware is needed to try it — the Mac’s own microphone works.
macOS · thirteen languages · nothing leaves the machine
Talkback you can read.
HUUTO turns the talkback channels on your console into a live text chat on your Mac. Every channel gets its own name and colour. Speech is recognised on the machine itself — no cloud, no account, no audio leaving the room.
MonitorsCan we get more vocal in wedge three21:04
StageGuitar DI is dead, patching the spare21:04
ProductionDoors in five, hold the soundcheck21:05
MonitorsdraftThat is better, thanks21:05
FOHCopy that — holding.21:05
The problem
Talkback is the channel you can't miss and can't hear.
Monitor world needs a word. Production is asking something mid-song. Someone at the patch wants to know which line went dead. It all arrives in the same ear, under a band at full tilt, in half-sentences and first names — and you catch maybe half of it.
HUUTO puts the same thing on screen. Not instead of the talkback bus, but alongside it, as a copy you can read.
How it works
Audio in, text out, nothing in between.
Any CoreAudio input
Dante Virtual Soundcard, MADI, AVB, SoundGrid, a USB mixer — or the Mac’s built-in microphone.
One channel per speaker
Each physical input channel is split out to mono and mapped to a person you name.
16 kHz mono
Resampled off the real-time thread, so the audio callback is never asked to do real work.
Voice activity detection
Analysis in 100 ms chunks with a 200 ms pre-roll, so the start of a word is not clipped off.
whisper.cpp
Runs on the GPU through Metal: one decoder state per channel, one shared model in memory.
Chat
An accepted line is written to disk first, and only then shown, exported and shared.
Speech is recognised on this Mac. Model files are downloaded once and checked against an exact file size, the GGML header and a SHA-256 before they are installed.
What it does
Built for the desk you are already standing at.
A draft before the transcript
On-device macOS recognition puts provisional text on screen in 0.54–0.91 s, measured, while Whisper works on the accurate version. Neither path sends audio anywhere.
Channels, named and grouped
Map each input to a person and a colour, then group them into Monitors, Stage, Production. A collapsed group still shows who is talking — hiding live speech was never an option.
Share it to the room
Anyone on the same network can open the transcript on another Mac over Bonjour, and type back to FOH from there.
Keywords, banners and OSC
Get a banner when your cue word is spoken — and fire an OSC message to the lighting or media desk from the same hit, so what you see and what the desk gets cannot disagree.
Sessions that survive
Every accepted line is written to disk the moment it is accepted. Search, filter by speaker, bookmark, and export to TXT or a multi-page PDF.
Languages
Language is a setting, not the product.
Pick the language in settings and HUUTO locks it when the session starts, so the decoder, the vocabulary and the hallucination filters stay in step for the whole show.
- English
- Finnish
- Swedish
- Norwegian
- Danish
- German
- Dutch
- French
- Spanish
- Italian
- Portuguese
- Polish
- Estonian
The engine is Whisper, which knows 99 languages. These are the ones HUUTO offers: short, overlapping, half-shouted talkback is a harder problem than clean dictation, and the tail of that list does not survive it.
Finnish also has dedicated fine-tuned models (Finnish-NLP) on top of the stock ones, which is unusual for a language this size. A strength, not the premise.
Local by default
The audio never leaves the Mac.
- No cloud service, no account, no API key. Recognition runs on this machine, on the GPU.
- No audio files are stored. Only the accepted text is written to disk.
- The instant draft uses on-device macOS recognition with requiresOnDeviceRecognition set, without exception. If the local asset is missing, the draft is switched off rather than quietly sent to Apple’s servers.
- Model files are downloaded once from Hugging Face and verified by exact size, GGML header and SHA-256 before they are installed.
Download
Coming soon.
HUUTO is distributed straight from this page — signed with a Developer ID and notarised by Apple, so it opens like any other app. No App Store, no account, no licence server.
What you will need
- macOS 14 or later
- Apple Silicon. Intel Macs have not been validated.
- Any CoreAudio input device — or none at all, if you just want to try it on the built-in microphone.
- Disk space for a speech model: 0.5–3.1 GB depending on which one you pick.
The app itself is a few megabytes. The speech model is a separate download on first run, so get it done before you are on a venue’s network.
Honest limits
What it does not do.
It is not intercom
HUUTO does not replace talkback, intercom, or any other safety-critical communication. Keep the bus you already have.
Transcription is probabilistic
Names, numbers, call signs and technical terms will sometimes come out wrong. Read it as a good copy, not a court record.
16 channels is a cap, not a promise
Sixteen is a safety limit on configuration. The reliable real-time maximum has not been benchmarked on any particular Mac, model or channel count.
It does not know who is speaking
The name comes from the channel you mapped by hand. This is not diarisation, and it cannot separate two people sharing one microphone.
LAN sharing is unencrypted
No TLS, no password, no verified identity. Anyone who can reach the network can find the service, read it and write to it. Use it on a trusted, isolated production network and nowhere else.