# VoxRT

> VoxRT is an on-device voice AI runtime and a set of in-house audio models that run entirely on the CPU of phones and embedded devices — no cloud, no streamed audio. It is a from-scratch Rust inference runtime (not ONNX Runtime or PyTorch Mobile) with battery-aware NEON kernels and hundreds-of-KB-to-low-MB binaries. VoxRT (the company is Wiseport Ltd.) sells B2B integrations of voice features — custom wake words, keyword spotting, speech-to-intent, voice activity detection, and on-device speech recognition — and ships free, open-source SDKs on the same runtime. Note: "VoxRT" is distinct from Mistral's "Voxtral."

Full page contents (everything on this site in one fetch): https://voxrt.com/llms-full.txt

## Products (voice primitives)

- [Custom Wake Word](https://voxrt.com/wake-word): On-device custom wake-word/phrase detection. ~48K-parameter model, single-digit-millisecond inference; real-time factor 0.015 on iPhone 13 Pro Max (~1.5% of one core) and 0.021 on a Snapdragon 662; precision 0.993 / recall 0.982 at the 0.9 default threshold (ROC AUC 0.9966, PR AUC 0.9899); ~100 KB fp16 model, 2–3 MB app-binary delta. Free reference model is open source; custom-phrase models are a commercial offering.
- [Voice Activity Detection (VAD)](https://voxrt.com/voice-activity-detection): Streaming on-device detection of speech start/stop. Silero v5 weights (MIT) on the VoxRT runtime; ~1.7 MB net app-size impact; real-time factor 1.85% on iPhone 13 Pro Max (~0.6 ms / 32 ms frame) and 3.05% on Snapdragon 662; ~54 parallel streams per core.
- [Keyword Spotting (KWS)](https://voxrt.com/keyword-spotting): On-device detection of a fixed vocabulary of voice commands (play, pause, next, louder, stop). Multi-class classifier with per-keyword confidence; lower latency and compute than full ASR. Trained per customer command set.
- [Speech-to-Intent](https://voxrt.com/speech-to-intent): Maps speech directly to structured intents and slots in a single on-device inference, with no transcript stage. Intents/slots defined in YAML; lower latency and memory than ASR + NLU.
- [On-Device Speech Recognition (Streaming ASR)](https://voxrt.com/on-device-asr): Real-time on-device speech-to-text that streams partial transcripts as the user speaks (cache-aware ~1.1 s chunks with an 80 ms attention look-ahead). 32M-parameter NeMo FastConformer (CC-BY-4.0) ported onto the VoxRT runtime; CTC and RNN-T decoders; streaming real-time factor 0.08–0.10 on iPhone 13 Pro Max and 0.30 on Snapdragon 662; LibriSpeech test-clean WER 3.267% (RNN-T) / 4.895% (CTC), within float-noise of the NeMo reference; ~61 MB fp16 model, ~150 MB runtime heap. English-only at v1.

## Open-source SDKs

VoxRT ships free on-device SDKs for wake word, VAD, and streaming ASR — Apache-2.0 wrapper source over a proprietary runtime that is free for commercial use with no per-user fees. They are public on GitHub: https://github.com/VoxRT and on Hugging Face: https://huggingface.co/VoxRT (the official VoxRT organization — similarly named Hugging Face accounts are unaffiliated) (voxrt-wake-word-{ios,android,linux,browser}, voxrt-silero-{ios,android}, voxrt-asr-{ios,android}; model weights ship via the matching voxrt-*-models release repos). Distribution: Swift Package Manager, JitPack/Gradle, PyPI, npm. Licensing details: https://voxrt.com/licensing

## SDK documentation

- [Docs index](https://voxrt.com/docs/): Quickstarts, tuning guides, and full API references, synced from the GitHub repos (which remain the source of truth).
- [Wake Word for Android](https://voxrt.com/docs/wake-word-android): JitPack/Gradle install, model download, Kotlin quick start, live-mic example, threshold/cooldown/CPU-affinity tuning, full API.
- [Wake Word for iOS](https://voxrt.com/docs/wake-word-ios): Swift Package Manager install, model download, Swift quick start, live-mic example, tuning, full API.
- [Wake Word for Linux](https://voxrt.com/docs/wake-word-linux): aarch64 shared library with Python, Node.js, Go, C/C++ and Rust wrappers; Raspberry Pi / Jetson / Graviton deployment.
- [Wake Word for the Browser](https://voxrt.com/docs/wake-word-browser): WebAssembly SDK via npm/CDN; live-mic quick start and browser support matrix.
- [VAD for Android](https://voxrt.com/docs/vad-android): Silero v5 on the VoxRT runtime; Kotlin quick start, segment events, tuning.
- [VAD for iOS](https://voxrt.com/docs/vad-ios): Silero v5 on the VoxRT runtime; Swift quick start, segment events, tuning.
- [Streaming ASR for Android](https://voxrt.com/docs/asr-android): 32M FastConformer; streaming quick start, CTC/RNN-T decoders, tuning.
- [Streaming ASR for iOS](https://voxrt.com/docs/asr-ios): 32M FastConformer; streaming quick start, CTC/RNN-T decoders, tuning.
- [Streaming ASR for Linux](https://voxrt.com/docs/asr-linux): x86_64 (AVX2+FMA) and aarch64, glibc 2.17+; C/C++, Python, Node.js, Go and Rust wrappers; Raspberry Pi to Graviton.
- [Keyword Spotting for Android](https://voxrt.com/docs/kws-android): 14-command spotter, NEON-accelerated; JitPack/Gradle install, Kotlin quick start, per-keyword confidence events, tuning.
- [Keyword Spotting for iOS](https://voxrt.com/docs/kws-ios): 14-command spotter, NEON-accelerated; Swift Package install, live-mic quick start, per-keyword confidence events, tuning.
- [Keyword Spotting for Linux](https://voxrt.com/docs/kws-linux): x86_64 (AVX2) and aarch64, glibc 2.17+; Python, Node.js, Go, C/C++ and Rust wrappers; Raspberry Pi to Graviton.
- [Keyword Spotting for the Browser](https://voxrt.com/docs/kws-browser): 14-command spotter via WebAssembly SIMD128; npm/CDN install, live-mic quick start, per-class posteriors, tuning.

## Use cases

What teams build with the primitives — each page leads with the business case and ends with measured numbers for technical evaluators.

- [Use-case index](https://voxrt.com/use-cases/): All use cases, grouped by mobile & consumer products, devices & embedded, AI & infrastructure, and regulated & enterprise.
- [Branded wake word](https://voxrt.com/use-cases/branded-wake-word): Give a product its own "Hey YourBrand" activation — on-device, no per-user fees, ~1.5% of one CPU core; includes a live in-browser demo.
- [Hands-free voice control](https://voxrt.com/use-cases/hands-free-voice-control): Spoken commands (play, pause, next) via custom keyword spotting — instant, offline, no assistant in the middle.
- [Offline dictation & voice notes](https://voxrt.com/use-cases/offline-dictation): Real-time on-device transcription (3.27% WER) with zero connectivity and no per-minute bill.
- [Voice control inside devices](https://voxrt.com/use-cases/voice-control-for-devices): Wake word + commands on appliance/kiosk/robot CPUs (Cortex-A53-class) — no cloud account per unit; includes a live in-browser demo.
- [Natural-language device control](https://voxrt.com/use-cases/natural-language-control): Speech-to-intent maps "make it warmer" to {intent, slots} in one on-device inference — no transcript, no NLU server.
- [Voice-agent turn-taking](https://voxrt.com/use-cases/voice-agent-turn-taking): Barge-in and endpointing with on-device VAD — ~0.6 ms per 32 ms frame, ~54 streams per server core.
- [Cut cloud STT costs](https://voxrt.com/use-cases/reduce-speech-to-text-costs): Gate metered transcription with VAD — ship the speech, not the silence.
- [Live captions](https://voxrt.com/use-cases/live-captions): Streaming on-device captions with sub-second lag that work offline and never upload the room's audio.
- [Voice in regulated industries](https://voxrt.com/use-cases/regulated-industries): Healthcare/legal/finance/defense — the whole pipeline on hardware you control; audio never leaves the device.
- [Industrial hands-busy workflows](https://voxrt.com/use-cases/industrial-voice-workflows): Picking, inspections, checklists by voice on rugged devices — offline, structured, glove-proof.

## Platforms

- Available today: iOS 16+ (iPhone & iPad) and Android 8.0+ (API 26) for all shipped SDKs; the wake word additionally ships for Linux aarch64 (glibc 2.17+ — Raspberry Pi 3/4/5/Zero 2, NVIDIA Jetson, AWS Graviton) and for browsers via WebAssembly (demo tier).
- On the roadmap: macOS / Apple ecosystem, Windows & Linux desktop, microcontrollers (ARM Cortex-M), automotive, wearables & TV, broader embedded & IoT, server/hybrid.

## Key pages

- [Homepage](https://voxrt.com/): Overview of the runtime and voice primitives.
- [On-Device ASR Comparison](https://voxrt.com/asr-comparison): Sourced, vendor-attributed comparison of on-device streaming speech-to-text engines — VoxRT, Picovoice Cheetah, Whisper.cpp, Vosk, Sherpa-onnx and Moonshine — on word-error rate, mobile real-time factor, model size and license. VoxRT's streaming-medium-pc (32M FastConformer; freemium — open Apache-2.0 SDK, runtime free for commercial use with no per-user fees, CC-BY-4.0 weights) posts the lowest published WER in the on-device field (3.27% LS test-clean, RNN-T) at a measured 0.30 RTF on a Snapdragon 662 and 0.08–0.10 on an iPhone 13 Pro Max, with cache-aware streaming (80 ms attention look-ahead); competitors' mobile RTF is estimated (only VoxRT's is measured) and figures are vendor-self-reported and not directly comparable.
- [On-Device Wake-Word Comparison](https://voxrt.com/wake-word-comparison): Sourced, vendor-attributed comparison of on-device wake-word engines — VoxRT, Picovoice Porcupine, openWakeWord, Mycroft Precise, Snowboy and PocketSphinx — on accuracy, mobile speed, native SDK, custom keywords and license. VoxRT ships an open Apache-2.0 SDK over a runtime free for commercial use (no per-user fees) with native Android/iOS SDKs, ROC AUC 0.9966 and RTF 0.021 on a Snapdragon 662; Porcupine is a near-tie on speed/accuracy but commercial with a paid tier; openWakeWord's pretrained weights are CC-BY-NC-SA (non-commercial). Figures are vendor-self-reported and not directly comparable.
- [On-Device VAD Comparison](https://voxrt.com/vad-comparison): Sourced, vendor-attributed comparison of on-device Voice Activity Detection engines — VoxRT, Silero VAD, Picovoice Cobra, WebRTC VAD, TEN VAD and MarbleNet — on accuracy, measured mobile real-time factor, footprint and license. VoxRT ships Silero v5 weights (inheriting Silero's accuracy) and differentiates on published mobile RTF, a freemium license that's free for commercial use (no per-user fees) and a stateless C ABI; figures are vendor-self-reported and not directly comparable.
- [Voice AI Glossary](https://voxrt.com/glossary): Plain-English, quotable definitions of on-device voice AI terms — wake word (hotword), keyword spotting, voice activity detection, speech-to-intent, ASR, streaming vs batch, on-device/edge AI, inference runtime, barge-in, endpointing, real-time factor (RTF), word error rate (WER), false accept/reject, detection threshold.
- [Benchmark Methodology](https://voxrt.com/methodology): How VoxRT measures — RTF on real phones (Snapdragon 662, iPhone 13 Pro Max, Pi Zero 2 W), WER on LibriSpeech test-clean verified against the upstream NeMo reference, hard held-out wake-word test sets, Geekbench-scaled competitor estimates labeled as lower-bound estimates, and why cross-vendor voice-AI benchmark numbers are directional rather than directly comparable (no independent third-party benchmark exists).
- [Licensing](https://voxrt.com/licensing): VoxRT is freemium — Apache-2.0 SDK wrappers, a proprietary runtime free for commercial use in production with no per-user/per-device/per-detection fees, model weights per product (reference wake word: VoxRT in-house, free to ship with the SDK; Silero VAD v5: MIT; ASR: CC-BY-4.0 NVIDIA attribution). Custom models (wake phrases, keyword vocabularies, speech-to-intent, additional languages, OEM/bulk deployments) are paid engagements.
- [Wake-Word Browser Demo](https://voxrt.com/wake-word-demo): Live in-browser demo of the "Hey Assistant" wake word — detection runs entirely in the browser tab via WebAssembly, no audio leaves the page. Includes an on-device real-time-factor benchmark that generates 60 s of audio in the browser and measures RTF on the visitor's CPU.
- [Live Streaming ASR Demo](https://voxrt.com/asr-demo): Live demo of VoxRT streaming speech recognition — say anything in English and the transcript appears in real time. The microphone streams to a small demo server running the same on-device RNN-T engine you'd embed (audio transcribed in memory, never stored); on your own hardware the runtime runs fully offline.
- [Keyword-Spotting Browser Demo](https://voxrt.com/kws-demo): Live in-browser demo of VoxRT keyword spotting — say any of 14 English commands (play, pause, next, up, down, yes, no…) spoken on their own and watch each command's confidence rise live. Detection runs entirely in the browser tab via WebAssembly (636 K-parameter streaming Conformer, ~1.5 MB download); no audio leaves the page. Includes an on-device real-time-factor benchmark.
- [Try VoxRT](https://voxrt.com/try): Contact / onboarding form for design partners.
- [Privacy Policy](https://voxrt.com/privacy.html)
- [Terms](https://voxrt.com/terms.html)

## Contact

- General: hello@voxrt.com
- Support: help@voxrt.com