LT_ausys — Lithuanian ears

Offline Lithuanian speech recognition, at home

A voice assistant needs a mouth and ears. The ears are speech-to-text (STT, also called ASR): they hear “Įjunk šviesą svetainėje” (“turn on the living-room light”) and turn it into text Home Assistant can act on. LT_ausys is Kristijonas Jakubsonas's Lithuanian model, which we converted to ONNX and compressed so it runs fast on a CPU with no GPU, served over the Wyoming protocol — the same one Home Assistant uses for its own voice services.

Why not Whisper

Whisper is the most popular open speech-recognition model, but it encodes a fixed 30-second window for every phrase, however short. On a CPU that is a cost of several seconds per command. LT_ausys is built on NVIDIA's Parakeet-TDT transducer, which processes only as much audio as it is given. Same machine (Ryzen 7 8845HS mini PC, 4 threads), same 127 home commands:

Measured on 2026-10-24 on one machine. Full tables, test sets and caveats: docs/results.md.

Getting started: Home Assistant in Lithuanian

Three ways in — pick the one closest to what you already run.

A. Python only

pip install "onnx-asr[cpu,hub]"

import onnx_asr
model = onnx_asr.load_model("RobertasTa/lt-ausys", quantization="int8")
print(model.recognize("command.wav"))   # 16 kHz mono WAV

B. A small Wyoming STT server of its own (Linux or Docker)

About 5 minutes: a Python environment with a systemd unit, or a single Docker image. The full recipe is in the README. The server listens on port 10302.

C. Already running wyoming-faster-whisper?

Since version 3.9 it has an onnx-asr engine, so no new server is needed:

wyoming-faster-whisper --uri tcp://0.0.0.0:10302 --stt-library onnx-asr \
    --model RobertasTa/lt-ausys --language lt --data-dir /data --download-dir /data

This route loads the larger fp32 model (2.5 GB) and ignores the thread count; for int8 use route B. Why.

Then Home Assistant

Settings → Devices & services → Add integration → Wyoming Protocol → the server's address and port 10302. In your Assist pipeline choose lt_ausys as the Speech-to-text engine and Lithuanian as the language.

The whole Lithuanian voice pipeline

For Home Assistant to understand and answer in Lithuanian you need three parts. All are free and run at home:

1. Ears — this project: Lithuanian speech recognition.
2. Mouth — a Lithuanian Piper voice, Reginutė or Ingutė: spoken answers with correct stress.
3. A microphone in the room — Home Assistant Voice PE, an ESPHome satellite, or a USB conference speakerphone via Speakerphone Assist.

Home Assistant already has Lithuanian intents (“įjunk šviesą svetainėje”).

Whose model this is

What it does not do

• It needs a wake word and voice activity detection (VAD) in front of it. The server transcribes whatever it receives and makes up text from pure noise — like every speech-recognition model.
• Home Assistant's Whisper add-on does not offer it by itself. Our proposal to pick it automatically for Lithuanian is still awaiting review; until then use route B or C above.
• The numbers come from one machine. Speed will differ on another CPU; int8 accuracy matched across the CPUs we tried.

Questions people ask

Can Home Assistant understand Lithuanian?

Yes. Home Assistant already has Lithuanian commands; what was missing is good speech recognition. The default Whisper base gets 88 % of Lithuanian words wrong, LT_ausys 10.6 %.

Does it need the internet or a GPU?

No. The model runs on a CPU (4 cores, 4 GB of RAM) and no audio leaves your network.

How is it different from Paprika?

Paprika is also Kristijonas Jakubsonas's model, but a Whisper one: accurate, yet several seconds per command on a CPU. LT_ausys is a different kind of model (Parakeet) and more than 15 times faster.

Can I use it in a commercial product?

The model's CC BY 4.0 licence allows it with attribution. Read the licences before you ship it.

Stuck? Ask an AI