Kea-T · running live in your browser

Speech intelligence you can take into the wilderness.

Specialized speech models engineered for tight power and memory budgets. Recognition runs entirely on your hardware.

Try Kea-T Request access
01

Kea-T, live.
600M parameters. 181 MB.
Running on your GPU.

Named for the kea, New Zealand's alpine parrot: clever, hardy and at home where few birds survive. Our models share that ethos, built small to work in hard places.

Derived from NVIDIA's Parakeet-TDT 0.6B v2, with 98% of its weights constrained to ternary values {−1, 0, +1}. Your audio never leaves this tab.

  • 181 MBweights, vs 2,472 MB original
  • −23%latency vs original, RTX 5090
  • 6.84%mean WER, vs 6.45% original
Model card on Hugging Face
Kea-T0.6B · ternary · WebGPU Not loaded

Downloaded once from Hugging Face, verified, then cached in your browser.

  1. 01Download181 MB · Hugging Face
  2. 02VerifySHA-256
  3. 03CompileWebGPU kernels
Audio stays in this tab rajb3/kea-t · CC-BY-4.0

02

Context shapes the model.
Vocabulary, acoustics and device limits.

Domain training

Specialized terminology and task-specific examples teach each model the distinctions its users depend on.

domain

Field acoustics

Wind, rotor wash, radio traffic. We train around the sounds that compete with a user's voice.

field

Device budgets

Compute, memory and battery limits guide model size and architecture from the start.

hardware
03

Efficiency at every layer.
Architecture, precision and runtime designed together.

FastConformer-TDT

A 24-layer FastConformer encoder with token-and-duration decoding, which skips ahead over frames instead of stepping through every one.

arch

GPU-native kernels

Custom compute kernels read the packed 2-bit weights directly: on datacenter GPUs, on laptops, and in this page through WebGPU.

kernels

Ternary weights

Every large encoder matrix holds only −1, 0 or +1, with one scale per row: 181 MB on disk against 2,472 MB for the original, at 6.84% mean WER versus 6.45%.

quant

Local deployment

Model files are checked against a published SHA-256 before they run. Audio stays on your device, with no telemetry or cloud dependency.

integrity
04

Corpus
Speech AI for casualty care.

in development

Hands on care.

A voice-operated, local-only guide to the Tactical Combat Casualty Care manual, with source-cited responses for trained responders.

Explore Corpus

Build with us.

We're seeking partners for early model evaluations.

Requests open soon. Your email has not been submitted.