Introducing Kea-B 0.1

We train the best small speech-to-text models in the world.

  • 91 MB
  • 4.3% WER on LibriSpeech test-other*
  • runs in a browser tab

We fine-tune Kea for your field: your vocabulary, your acoustics and your device budget.

Fine-tune Kea How Kea compares

01

The Kea family.
Pushing the Pareto frontier for size and accuracy.*

Kea is our family of small, English-only speech-to-text models. They are named for New Zealand's alpine parrot, a small and famously clever bird.

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}. Kea-B 0.1, the current release, approaches the accuracy of Parakeet v2 at 1/25 its size.

Kea-B 0.1 on Hugging Face Kea-T on Hugging Face
  • 91 MBKea-B 0.1 weights, vs 181 MB for Kea-T and 2,472 MB for Parakeet v2
  • 6.77%Kea-B 0.1 mean WER, vs 6.84% for Kea-T and 6.45% for Parakeet v2
  • −37%Kea-B 0.1 latency vs Parakeet v2, 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 0.1 91 MB at 4.3%.

02

Tuned for your field.
A model that recognizes the words your field uses.

Your vocabulary

Medicine, law, aviation and the trades each have their own terms. We fine-tune Kea on yours, so it gets right the words your work depends on.

vocabulary

Your acoustics

We train on audio that sounds like yours: your rooms, your microphones and the noise around your speakers.

acoustics

Your device budget

We size the model to the compute, memory and battery you have. It runs on your hardware, and the audio stays there.

hardware
03

Small by design.
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 1- and 2-bit weights directly: Kea-T's on datacenter GPUs and laptops, and both models' in this page through WebGPU.

kernels

1-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, and Kea-B 0.1 keeps its decoder in 8-bit integers: 91 MB and 181 MB on disk against 2,472 MB for the original.

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

Sorello
Free dictation that runs on your Mac.

free · macOS beta

Dictate into any text field.

Hold a shortcut, speak, and the text appears at your cursor. No account and no network transcription. Powered by Kea-B 0.1, entirely on your device.

Download for macOS Try the nightly build

Fine-tune Kea for your field.

Tell us about your vocabulary, your audio and the hardware it has to run on. Fine-tuning inquiries open soon.