Live captions that never send audio anywhere
Caption speech as it happens — in calls, classrooms, and venues — with a streaming recognizer that runs on the device doing the captioning. Captions keep working with no connection, and the conversation being captioned never leaves the hardware it was spoken into.
Why captions belong on the device
Accessibility without an asterisk
Captions that depend on connectivity fail exactly where they're needed — trains, basements, crowded venues, planes. On-device captioning works everywhere the product does, with no "requires internet" footnote.
Private by construction
Captioning a call, a lecture, or a meeting means processing everything everyone says. When recognition is local, that audio never transits a captioning vendor — a claim your privacy policy can make in one sentence.
Real-time, actually
Partial captions appear as words are spoken and finalize when the sentence ends — streaming with an 80 ms attention look-ahead, not a batch job pretending to be live.
Captions across products
Anywhere speech needs to be readable the moment it's spoken.
Calling & communicationIn-call captions
Live captions inside voice and video calls — including the calls too sensitive to route through a third-party captioner.
Events & venuesCaption displays on local boxes
A Linux box beside the stage drives the caption screen — the venue's audio never streams out of the building.
EducationLectures, captioned and kept
Classroom capture that produces readable text in real time and a transcript after — on school-owned hardware.
Deaf & hard-of-hearing toolsPersonal captioning
A phone on the table captions the conversation around it — offline, private, no account required.
Media & playersCaptions for uncaptioned content
Live caption of audio and video that shipped without subtitles, rendered on the viewer's own device.
Workplace softwareMeetings with a no-upload guarantee
Caption and transcribe internal meetings while keeping a promise security teams can verify: nothing leaves the laptop.
From audio to readable text
Add the SDK and model
One dependency brings the runtime and the 32M-parameter streaming FastConformer — English with punctuation and capitalization at v1.
Stream the audio source into the engine
Microphone, call audio, or system audio — partial transcripts stream out as words are spoken, with an 80 ms attention look-ahead.
Render partials, replace with finals
Show the partial line as it grows, then swap in the finalized sentence when the recognizer commits it on end of speech — the familiar live-caption feel.
Keep or discard the transcript
The finalized text is yours: display-only for pure captioning, or stored as a transcript — either way it never had to leave the device.
The numbers behind the promise
Accuracy and streaming speed are measured and published — see the benchmark methodology and the on-device ASR comparison.
Pipeline. Audio source → on-device VAD → streaming FastConformer (CTC/RNN-T) → partial captions while speaking, finalized lines on end of speech. At a streaming real-time factor of 0.08–0.10 on an iPhone 13 Pro Max, roughly 90% of the core remains free for the rest of the product.
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 or per-minute fees; domain or language tuning is a paid engagement — see licensing.
Live captions, answered
How far do captions lag behind the speaker?
The recognizer streams with an 80 ms attention look-ahead and processes a 1.12 s chunk in about 90 ms on an iPhone 13 Pro Max, so partial captions trail the voice by well under a second and finalize as each sentence ends.
How accurate are the captions?
3.27% word-error rate on LibriSpeech test-clean with the RNN-T decoder — within floating-point noise of the upstream NVIDIA NeMo reference. Real-world accuracy depends on audio quality and domain; the methodology page explains how the numbers are produced.
What languages can be captioned?
English at v1, with punctuation and capitalization. Additional languages are roadmap items driven by customer demand.
Does any of the captioned audio leave the device?
No — recognition is entirely local. That's the point: captioning everything everyone says is only acceptable when nothing anyone says is uploaded.