Your wake word. Fully on‑device.
Type your English wake word, hear it in 10 synthetic voices and review the near-misses it should ignore. Then download a custom model for VoxRT's free SDK, with no sales call.
- No per-device, per-user or per-detection fees
- Single-digit-millisecond inference, even on budget Android
Example detection score
Start free. Swap in your own wake word when you're ready.
-
1
Install for free
The SDK and the "Hey Assistant" model cost nothing to install. Get them on GitHub
-
from_path("voxrt_wake_word.vxrt")2Build and test
Build your product with "Hey Assistant" and check it does what you need.
-
from_path("hey_vox.vxrt")3Swap one file
Create your wake word in VoxRT Studio, download the file and change this one line. The rest of your code stays the same.
A wake word that never leaves the device
Your wake word, your brand
Single words, brand names or multi-word activations, trained for your exact wake word.
Trained against near-misses
Hard negatives built from your wake word (partial phrases, rhymes, similar openings), plus augmentation for background noise, distance from the mic and a wide range of accents.
Fully on-device
No microphone audio ever leaves the device. Works with no network at all.
Battery-aware
Gated by on-device voice activity detection, so it idles when nobody is speaking.
Tunable threshold
Trade precision for recall with one confidence setting.
Runs where you ship
iOS 16+, Android 8+ and Linux aarch64 (Raspberry Pi, Jetson), plus a browser build for demos.
Start demo and say "Hey Assistant"
Your audio never leaves this page. The model reacts to "Hey Assistant" and ignores all other speech, including similar-sounding phrases like "Hey sister" or "Assist me".
"Hey Assistant"
High accuracy in a ~100 KB model
Benchmarked on our "Hey Assistant" reference model. Results vary by wake word.
- Precision
- 99.3%
- precision at 98.2% recall (default 0.9 threshold)
- ROC AUC
- 0.9966
- ROC AUC on 11,656 held-out test clips
- False-positive rate
- 0.5%
- false-positive rate: 34 of 6,416 hard negatives (default threshold)
- Model size
- ~100 KB
- model file; ~600 KB total Android APK impact
Test set: 5,240 positive utterances and 6,416 hard negatives (partial phrases, other assistants' wake words, phonetic neighbours, speech and noise), speakers disjoint from training. Compare wake word engines
Accuracy at every threshold
| Threshold | Precision | Recall | F1 | False-positive rate | False positives |
|---|---|---|---|---|---|
| 0.5 | 0.864 | 0.995 | 0.925 | 12.8% | 822 / 6,416 |
| 0.85 | 0.957 | 0.987 | 0.972 | 3.7% | 234 / 6,416 |
| 0.9 (default) | 0.993 | 0.982 | 0.987 | 0.5% | 34 / 6,416 |
| 0.95 | 0.997 | 0.769 | 0.868 | 0.2% | 12 / 6,416 |
Average precision 0.9899.
Start free, or train your own
Custom self-serve$199per wake word, one-time Your own English wake word, trained for you. |
|||
|---|---|---|---|
| Wake phrase | "Hey Assistant" | Your own phrase | Your own phrase |
| Language | English | English | English + other languages |
| Model size | ~100 KB | ~100 KB | ~100 KB |
| Offline and on device | Included | Included | Included |
| Near-misses built from your phrase | Not included | Included | Included |
| Retrains | Not included | Up to 3 included | Included |
| Trained for your domain, accents, devices | Not included | Not included | Included |
| Time to get it | Download now | Setup in under 10 minutesModel within 24 hours | Typically 5–10 business days |
Your own English wake word, trained for you.
- Wake phrase
- Your own phrase
- Language
- English
- Model size
- ~100 KB
- Near-misses built from your phrase
- Included
- Retrains
- Up to 3 included
- Time to get it
- Setup in under 10 minutes; model within 24 hours
- Wake phrase
- "Hey Assistant"
- Language
- English
- Model size
- ~100 KB
- Time to get it
- Download now
- Wake phrase
- Your own phrase
- Language
- English + other languages
- Trained for your domain, accents, devices
- Included
- Time to get it
- Typically 5–10 business days
In every option
- Accuracy
- 99.3% precision at 98.2% recall*
- Always-on CPU
- ~1.5% of one CPU core on an iPhone 13 Pro Max*
- Runs on
- iOS 16+ · Android 8+ · Linux aarch64 (Raspberry Pi, Jetson)
- Privacy
- Audio never leaves the device.
- License
- Commercial use. No per-device, per-user or per-detection fees. Licensing
*Measured on the "Hey Assistant" reference model. Results vary by wake word.
Questions about the options? Email [email protected].
-
1
Type your wake word
A distinctive English word or phrase; two to four words work best.
-
2
Check the pronunciation
Hear it in 10 synthetic voices and fix the spelling if needed.
-
3
Review similar phrases
See the rhymes and near-misses your model learns to ignore. Add your own.
-
4
Pay once, we train
Download your model when training finishes, within 24 hours.
Fast enough to leave always-on
Measured on real hardware with the "Hey Assistant" reference model. Results vary by wake word.
- iPhone real-time factor
- 0.015
- real-time factor on iPhone 13 Pro Max: ~150 µs per 10 ms frame
- CPU while listening
- ~1.5%
- of one CPU core during continuous listening, iPhone 13 Pro Max
- Budget Android real-time factor
- 0.021
- real-time factor on a Snapdragon 662 (2020 budget Android), about 2% of one core
- Parameters
- ~48K
- parameters in the model, small enough for microcontrollers
Footprint: from ~600 KB in your product
- Android net APK impact (arm64, incl. model)~600 KB
- iOS binary size increase after extraction and dead-code elimination2–3 MB
- Swift wrapper source (one file)~7 KB
- Native xcframework, compressed (device + simulator)~19 MB
- Wake word model (fp16)~100 KB
Where it runs
The same ~100 KB model runs behind native mobile SDKs and a Linux library. Audio is 16 kHz mono in, and nothing ever leaves the device.
iOS
- Minimum OS
- iOS 16.0
- CPU
- arm64 (device); arm64 + x86_64 (simulator)
- Install
- Swift Package Manager
Android
- Minimum OS
- Android 8.0 (API 26)
- CPU
- arm64-v8a (NEON); x86_64 (emulator)
- Install
- JitPack / Gradle
Linux
- Minimum OS
- glibc 2.17+: Ubuntu 18.04+, Debian 10+, RHEL 7+, Raspberry Pi OS
- CPU
- aarch64
- Install
- Shared library (.so) with Python, Node.js, Go, C/C++ and Rust wrappers
On Linux, it's validated on Raspberry Pi 3, 4, 5 and Zero 2 W, NVIDIA Jetson Nano and Orin Nano, AWS Graviton, and boards such as Rock Pi, Orange Pi and Khadas (Cortex-A53 / A55). It ships as a single ~480 KB shared library.
Custom models ship through the native SDKs. A WebAssembly build (@voxrt/wake-word-browser on npm) is also available for demos and prototypes with "Hey Assistant".
The runtime is the product
VoxRT is a from-scratch inference runtime for on-device speech models: no ONNX Runtime, no PyTorch Mobile, no LiteRT. It's a custom Rust core, sized and tuned for streaming voice on constrained, low-power hardware.
Custom Wake Word is one product on that runtime, alongside voice activity detection and streaming ASR. They share the same Rust runtime and NEON kernels: the runtime is the product, and the models are what it runs.
Custom wake words, answered
What exactly do I get?
A .vxrt model file for your wake word. Load it into the free VoxRT wake word SDK on iOS, Android or Linux and set your confidence threshold. The SDK takes 16 kHz mono audio.
How long does it take?
Under 10 minutes of your time to type the wake word, check the pronunciation and review similar phrases. Training starts when payment clears and finishes within 24 hours; we email you when it's ready.
How do I test my model?
Build your integration with the free "Hey Assistant" reference model that ships with the SDK (SDKs on GitHub), then drop in your model file in its place. Same API, same code: only the model file changes.
What if I want changes after training?
Up to 3 retrains are included. Email [email protected] with what you'd like to change.
Can I ship it in a commercial product?
Yes. You can ship your custom model in commercial products. The SDK and the "Hey Assistant" model are free for commercial use, and the SDK wrappers are open source under Apache-2.0. There are no per-device, per-user or per-detection fees. See licensing for the full picture.
How much battery does always-on listening use?
About 1.5% of one CPU core on an iPhone 13 Pro Max and about 2% on a 2020 budget Android phone (Snapdragon 662). The detector is gated by voice activity detection, so it idles when nobody is speaking.
Does it keep listening in the background or with the screen locked?
Yes, if you set up your product that way: always-on listening is what it's built for. Whether audio capture continues in the background is controlled by your code and the platform's background-audio settings.
Does audio go to the cloud?
No. Detection runs entirely on the device. Microphone audio never leaves it and no network connection is needed.
What platforms does it run on?
iOS 16+ and Android 8.0+ (API 26) as native Swift Package and Gradle/JitPack modules, plus Linux aarch64 (glibc 2.17+, such as Raspberry Pi, NVIDIA Jetson and AWS Graviton) with Python, Node.js, Go, C/C++ and Rust wrappers. A WebAssembly build is available for in-browser demos with "Hey Assistant".
Which wake words and languages are supported?
Self-serve takes English wake words. Distinctive words with clear syllables work best; two to four words is ideal, like "Hey Marina" or "OK Nova". Other companies' assistant wake words such as "Hey Siri" or "Alexa" are rejected. For other languages, talk to us about a tailored model.