Dictation that works at 30,000 feet and three floors underground
Field crews, clinicians, inspectors, travelers — the people who most need to capture notes by voice are often the ones without a signal. VoxRT transcribes speech in real time on the device itself: no connection required, nothing uploaded, and no per-minute transcription bill.
Why teams move dictation on-device
Capture happens where work happens
Basements, job sites, aircraft, rural routes, hospital wings with dead spots — dictation keeps working because there is nothing to connect to. Sync the text later, or never.
No per-minute meter running
Cloud transcription bills by the audio minute, so your heaviest users are your biggest cost. On-device recognition costs the same at minute one and minute one million: nothing.
Notes stay private by default
Dictation is where the sensitive material lives — patient details, client matters, unreleased plans. When transcription happens on the device, none of it transits a third party.
Who dictates offline
Anywhere the note matters more than the signal.
Field service & inspectionsFindings narrated on-site
Inspectors walk the site and talk; the report text is on the device before they reach the truck — coverage or not.
HealthcareNotes at the bedside
Clinical documentation captured where care happens, with audio that never leaves the device — a shorter compliance conversation.
Sales & CRMThe parking-garage debrief
Reps dictate meeting notes in the elevator, the garage, the plane — the CRM entry syncs when the signal returns.
Journalism & researchInterviews without uploads
Transcription on the reporter's own machine — sources' words never sit on a transcription vendor's servers.
Logistics & deliveryNotes from the cab
Proof-of-delivery and exception notes spoken in the cab on rural routes, transcribed with zero connectivity.
Personal productivityJournals & idea capture
Voice-first note products where instant, local transcription is the feature — no account, no upload, no wait.
From microphone to text
Add the SDK and model
One dependency brings the runtime and the 32M-parameter FastConformer recognizer. It ships with English, including punctuation and capitalization.
Stream the microphone into the engine
Partial transcripts appear as the user speaks — the model streams with an 80 ms attention look-ahead, so text trails the voice by well under a second.
Finalize on end of speech
On-device voice activity detection endpoints each utterance and the recognizer finalizes it — clean sentences, no manual stop button required.
Store and sync on your terms
The output is plain text in your product. Save it locally, sync it when there's signal, or feed it to whatever comes next — that part never involves VoxRT.
The numbers behind the promise
Accuracy and speed below are measured and published — see the benchmark methodology and the on-device ASR comparison.
Pipeline. Microphone → on-device VAD (endpointing) → streaming FastConformer with CTC and RNN-T decoders → partial transcripts while speaking, finalized text on pause. Cache-aware streaming with an 80 ms attention look-ahead keeps latency conversational.
Platforms. iOS 16+ (Swift Package Manager) and Android 8.0+/API 26 (JitPack/Gradle), plus Linux aarch64 for device-class hardware — validated on Raspberry Pi 3/4/5 and Pi Zero 2 W, NVIDIA Jetson Nano and Orin Nano, AWS Graviton and similar Cortex-A53/A55 boards — shipped as a single ~480 KB shared library with Python, Node.js, Go, C/C++ and Rust wrappers. Quickstarts: iOS · Android.
Licensing. The published English model and SDK are free for commercial use with no per-user, per-device, or per-minute fees. Custom vocabulary, domain, or language work is a paid engagement — see licensing.
Offline dictation, answered
How accurate is on-device transcription, really?
3.27% word-error rate on LibriSpeech test-clean with the RNN-T decoder — within floating-point noise of the upstream NVIDIA NeMo reference and the lowest published WER in the on-device field VoxRT has surveyed. Method and comparisons are on the methodology page.
Does it really work with no connection at all?
Yes — airplane mode is the easiest demo. The model, the decoder, and the endpointing all live on the device; connectivity only matters if your product chooses to sync the resulting text somewhere.
What languages are supported?
English at v1, with punctuation and capitalization included. Additional languages are on the roadmap and customer-driven — if you need one, that's a conversation worth having early.
What does it cost per minute?
Nothing — there is no metering. The published model is free for commercial use; the economics are one integration, not a usage bill. See licensing.