Domain training
Specialized terminology and task-specific examples teach each model the distinctions its users depend on.
domain
Kea-B · 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.
Runs inside this browser tab.
The model is 105 MB. Best on Wi-Fi and a laptop or desktop.
Downloading weights
0 / 105 MB
huggingface.co/rajb3/kea-b
Record
0:00 / 0:30
30-second limit
Transcript
This browser doesn't expose WebGPU, so Kea-B can't run here. Try a recent Chrome, Edge or Safari on a laptop or desktop.
Kea-B transcripts of the sample recordings
- 01Download105 MB · Hugging Face
- 02VerifySHA-256
- 03CompileWebGPU kernels
Or try a recording:
Audio stays in this tab
rajb3/kea-b · CC-BY-4.0
01
The Kea family.
The Kea family.
Pushing the Pareto frontier for size and accuracy.*
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.
Both models derive from NVIDIA's Parakeet-TDT 0.6B v2, with 98% of their weights constrained to a handful of values. Kea-B's are binary {−1, +1}, one bit each; Kea-T's are ternary {−1, 0, +1}. Your audio never leaves this tab.
- 105 MBKea-B weights, vs 181 MB for Kea-T and 2,472 MB original
- 7.05%Kea-B mean WER, vs 6.84% for Kea-T and 6.45% original
- −23%Kea-T latency vs original, RTX 5090**
* Via the Open ASR Leaderboard (October 2026), WER on LibriSpeech test-other: Parakeet-TDT 0.6B v3 is
714 MB (8-bit GGUF) at 3.1%; Niagara-9m is 140 MB at 14.5%. Kea-T is 181 MB at 4.2%, and Kea-B
105 MB at 4.6%.
** Kea-B latency benchmarks pending.
02
Context shapes the model.
Context shapes the model.
Vocabulary, acoustics and device limits.
Field acoustics
Wind, rotor wash, radio traffic. We train around the sounds that compete with a user's voice.
fieldDevice budgets
Compute, memory and battery limits guide model size and architecture from the start.
hardware
03
Efficiency at every layer.
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.
archGPU-native kernels
Custom compute kernels read the packed 1- and 2-bit weights directly: Kea-T's on datacenter GPUs and laptops, and both models' in this page through WebGPU.
kernels1-bit and ternary weights
Every large encoder matrix holds only −1 or +1 (Kea-B) or −1, 0 or +1 (Kea-T), with one scale per row: 105 MB and 181 MB on disk against 2,472 MB for the original.
quantLocal 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
in development
Corpus
Speech AI for casualty care.
Hands on care.
A voice-operated, local-only guide to the Tactical Combat Casualty Care manual, with source-cited responses for trained responders.
Explore CorpusBuild with us.
We're seeking partners for early model evaluations.
Requests open soon. Your email has not been submitted.