Voice features where audio can never leave the device
Healthcare, legal, finance, defense: the fields with the most to gain from voice are the ones that can't ship audio to a cloud API. VoxRT's entire pipeline — wake word, voice activity, commands, intent, transcription — runs on hardware you control, so the data-flow answer is one sentence: the audio never leaves.
What on-device buys a compliance review
A one-sentence data-flow story
Vendor security reviews, DPAs, and audit questionnaires get dramatically shorter when no audio is transmitted to, processed by, or stored with a third party. There is no audio data flow to diagram.
Voice for the people who need it most
Clinicians, attorneys, and advisors are the ultimate hands-busy, documentation-burdened users — the ones voice was supposed to help before compliance took it off the table.
No new vendor in the audio path
The models run inside your product on your hardware. VoxRT never receives, processes, or stores your users' audio — by architecture, not by policy.
Voice inside the boundary
The same primitives, deployed where the data can't travel.
Clinical documentationDictation at the bedside
Notes captured where care happens, transcribed on the device — patient audio never in transit to a transcription vendor.
Patient rooms & care"Hey Nurse, I need help"
Hands-free assistance calls from bedside tablets and wearables, detected locally in always-listening mode.
LegalPrivileged dictation
Matter notes and memos dictated without placing privileged audio on any third party's infrastructure.
Financial servicesMeeting notes without processors
Advisor and client-meeting capture with no new data processor added to the audio path.
Government & defenseAir-gapped by design
The pipeline needs no network at all, so it runs identically in disconnected and classified environments.
Pharma & labsSterile, gloved, documented
Voice-driven protocol steps and observations where gloves and sterility rule out keyboards entirely.
Bringing voice inside the boundary
Draw the data-flow diagram — it's short
Audio enters the microphone; events and text come out of the SDK; nothing crosses the network boundary. That's the whole diagram reviewers need to see.
Pick the primitives for the workflow
Dictation (streaming ASR), hands-free commands (keyword spotting), structured data entry (speech-to-intent), always-listening activation (wake word) — all on one runtime.
Deploy on hardware you manage
Managed iOS and Android fleets, or Linux devices and terminals — including fully air-gapped environments, since nothing in the pipeline needs a connection.
Answer the review with published evidence
Accuracy and performance claims are published with reproducible methodology, so evaluators can verify numbers rather than take a vendor's word — see the methodology page.
The architecture, for evaluators
Every capability below runs on one CPU-only runtime with published, reproducible benchmarks — see the methodology page.
Pipeline. Microphone → on-device VAD → whichever primitives the workflow needs (wake word, keyword spotting, speech-to-intent, streaming ASR) → events and text delivered to your application locally. All models execute on the device CPU; there is no telemetry path for audio.
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.
Licensing. Published SDKs and models are free for commercial use; custom models (phrases, vocabularies, domain tuning) are one-time engagements — no per-seat or per-device fees to model in a procurement review. See licensing.
Regulated-industry questions, answered
Does using VoxRT make our product HIPAA or GDPR compliant?
No tool makes a product compliant by itself — but on-device processing removes the hardest part of the problem: there's no audio (and no PHI or personal data inside it) being transmitted to or processed by a third party, so that entire data flow and its paperwork disappear. Your counsel still owns the overall assessment.
Can it run fully air-gapped?
Yes. Models ship inside your product and nothing in the pipeline performs network calls, so disconnected and classified environments run the identical stack.
Does VoxRT ever see our audio, transcripts, or intents?
No. The SDK processes audio locally and returns results to your code; there is no path by which audio, text, or derived data reaches VoxRT.
Can we combine dictation, commands, and a wake word?
Yes — all five primitives share one runtime, so a bedside product can carry an always-listening help phrase, structured voice data entry, and full dictation as one dependency.