AT YOUR DESK01
Prompts without the performance
Explain the change, the error, or the edge case in a low voice. The rest of the room does not need to join the conversation.
BUILT FOR CODING IN SHARED SPACES
Voice your prompt.
Keep it to yourself.
Hold your iPhone close and press to talk. You can keep your voice down while your Mac turns it into text locally—right where your cursor already is.
01 / THE SIMPLE IDEA
Your Mac is across the desk. Your iPhone can be inches away.
A distant laptop microphone picks up your keyboard, ventilation, and everyone nearby along with your voice. Move the microphone to your mouth and your voice reaches it first. That is what lets you speak more softly without asking the model to rescue poor audio.
IPHONE → MAC / HOW IT FEELS
Speak softly into your phone. Watch the words land in Codex.
The iPhone handles close-range capture. Encrypted audio goes to the Mac you paired, where it is transcribed locally and inserted at the active cursor.
Building route preferences
ScenicRouteRanker.swiftBuild a walking-route map that ranks shade, noise, and slope instead of simply choosing the fastest path.
NEAR-FIELD AUDIO02
Give your voice the head start
Close-talk capture does not silence the office. It makes your voice stronger than the room before recognition even begins.
CODING FLOW03
Stay in the tools you already use
Keep focus in Codex, Claude Code, Xcode, Terminal, or a browser. The transcript arrives at the cursor.
02 / WHAT MATTERS AT A DESK
Useful numbers, not mystery metrics.
These are the facts that shape everyday use: how close the microphone is, where audio goes, and how much friction remains after you speak.
A microphone you can bring close
Your iPhone can sit inches from your mouth while your MacBook stays 50–80 cm away on the desk.
Less voice for the same input level
A distance-only estimate for 5–10 cm versus 50–80 cm. The room, device angle, and microphones affect real results.
No cloud transcription hop
Recognition runs on your Mac. Once the model is ready, local use does not need the public internet.
Nothing to clean up later
Audio is processed in memory and released. It is not left behind as a recording on either device.
The result goes where you were typing
No second app, clipboard shuffle, or context switch between speaking and coding.
Short local history by default
Mac transcript history defaults to one hour and can be turned off. It never syncs to an undervoice server.
The boundary you can verify:paired devices only; no audio upload or audio files; insertion stops when focus changes; password and secure-input fields are excluded.
03 / MODELS AND LANGUAGES
Seven core languages. English technical terms included.
undervoice uses language-specific packs instead of pretending one model is equally good at everything. Chinese, Japanese, German, French, Spanish, and Korean packs accept English library names, variables, and commands; a dedicated English mode is included.
中文 + EnglishEnglish日本語 + EnglishDeutsch + EnglishFrançais + EnglishEspañol + English한국어 + English
ZH + ENLOCAL
Chinese with English in the sentence
A bilingual streaming model provides the live preview; Paraformer or the accuracy-first FireRedASR2 CTC produces the final text. Punctuation is restored locally.
- Runtime
- sherpa-onnx
- Optimization
- INT8 ONNX
- DESIGNED FOR
- fast preview, stronger final pass
EN / DE / FR / ESLOCAL
Four European language packs
NVIDIA Parakeet TDT 0.6B v3 powers the product-tuned English, German, French, and Spanish packs.
- Runtime
- sherpa-onnx
- Optimization
- INT8 ONNX
- UPSTREAM BENCHMARK
- 6.34% average WER
JA + ENLOCAL
Japanese that keeps the English terms
Kotoba Whisper Bilingual v1.0 was trained for Japanese, English, and speech that moves between the two.
- Runtime
- whisper.cpp
- Optimization
- Q5_0
- UPSTREAM BENCHMARK
- 9.3–16.8% Japanese CER
KO + ENLOCAL
Korean with names, commands, and acronyms
Whisper Large v3 Turbo covers Korean speech and the English technical language inside it while keeping the local final pass practical.
- Runtime
- whisper.cpp
- Optimization
- Q5_0
- CHOSEN FOR
- multilingual coverage and local speed
Read the accuracy figures carefully
WER and CER are lower-is-better. The published 6.34% and 9.3–16.8% figures come from upstream full-model test sets—not an overall undervoice score for close-talk iPhone audio, quantized models, or mixed-language speech. Different datasets and metrics do not make a fair leaderboard.
04 / IPHONE VS. MACBOOK MICROPHONE
Fix the input before judging the model.
Both routes can use the same local model on your Mac. The difference is where the audio begins: inches from your mouth or across the desk.
| Shared-office use | iPhone held close | MacBook microphone |
|---|---|---|
| Mouth-to-mic distance | 5–10 cm | 50–80 cm |
| Speaking quietly | Your voice reaches the mic first | Voice, keyboard, HVAC, and nearby speech arrive together |
| Voice level needed | Stay at a low voice | Often requires speaking up at desk distance |
| Where recognition runs | Locally on your Mac | Locally on your Mac |
| Best fit | Open offices and shared desks | Private rooms or conversations you do not mind sharing |
The 14–24 dB estimate uses the free-field relationship 20 × log10(r₂/r₁) for 5–10 cm versus 50–80 cm. Reflections, placement, and speaking direction change real-world results.
05 / SECURITY AND NETWORKING
Your recording goes straight to the Mac you chose.
undervoice has no speech relay. On the same network, iPhone connects directly to Mac; across networks, you can use your own Tailscale setup. App-level authentication and encryption stay in place either way.
Installation and permission details →Only paired devices get in
Sharing Wi-Fi is not enough. A new iPhone must be explicitly paired or approved after a reset.
Encrypted before it leaves iPhone
Audio, transcripts, and control messages are authenticated and encrypted end to end.
A fresh key for every connection
Reconnect and a new session key is derived. Captured traffic from an old session cannot simply be replayed.
Audio never becomes a file
Recordings live in memory only. Diagnostics exclude audio and transcript text.
No undervoice speech cloud
On a local network, audio travels directly from iPhone to Mac. The model runs there.
Bring your own private network
For another Wi-Fi or cellular connection, use your own Tailscale network without dropping undervoice encryption.
06 / TYPE ANYWHERE
Wherever the cursor is, that is where the words go.
CodexClaude CodeXcodeTerminalAny text field
“Keep the public API, replace retries with exponential backoff, and add three boundary tests.”
listeningYour Mac remembers the active app and selection when dictation starts. If focus or selection changes, insertion pauses. Password fields and controls using Secure Keyboard Entry are never written to.
TALK TO YOUR MAC, NOT THE ROOM
Speak softly. Code freely.
Install the model on Mac, scan the pairing code with iPhone, then hold to talk whenever an idea is faster to say than type.