Reginutė
A Lithuanian voice that speaks on your own computer.
Lithuanian. Local. Free.
Full documentation in English: README_EN.md
A Lithuanian voice that speaks on your own computer.
Lithuanian. Local. Free.
Full documentation in English: README_EN.md
This is how she sounds. Synthesis, not a recording.
This is a synthetic voice. All audio samples are generated by AI; they are not recordings of a person. If you publish content voiced with it, check whether it needs an AI label (EU AI Act, Article 50, applicable since 2 August 2026).
The voice lives on your computer or home server. No account, no API key, no monthly fee, no character limit. Your text goes nowhere.
Lithuanian has three pitch accents, and standard espeak-ng puts the stress on the wrong syllable in about half of the words. So this voice does not use espeak's stress: it takes it from a dictionary of 189,262 word forms.
„Kompensacija nuo 2000 eurų“, „15:30“, „25 °C“ — numbers, times, units and abbreviations are expanded into words with the right Lithuanian endings before synthesis, not read digit by digit.
Connects over the Wyoming protocol on its own port, next to your existing Piper add-on — nothing that already works is touched. Tested live with a Voice PE speaker.
Latvian has been in the Piper catalogue since October 2024, Estonian since August 2026. Lithuanian never was — until 17 September 2026, when Reginutė became the first Lithuanian voice in the official catalogue. In the Home Assistant Piper add-on she is expected with Piper 1.9.0 (not released yet).
Code — GPL-3.0, voice — CC-BY-4.0, inherited from the corpus. Use it, change it and share it, with attribution. You download the whole model; nothing is rented.
The voice is trained on the synthesis part of the LIEPA corpus of Vilnius University — studio recordings of a professional actress (~3 h, CC-BY-4.0). These are the same recordings that the 2015 LIEPA synthesizer for blind people was built from; Reginutė is a digital daughter of that voice, not a competitor.
Everything you hear is still hers: her timbre, her pace, her way of ending a sentence. The model only learned to rearrange it.
You need the released piper-tts wheel — no fork, no patches. The voice also needs
the stress-aware phonemizer from the same folder: without it Piper would feed the model bare
letters and it would not speak Lithuanian.
Two paths, both tested on a clean Linux container. Home Assistant — a Wyoming server: its own Python environment, every file from Hugging Face and a systemd service, about 5 minutes — README_EN, “Install path B”. Plain Piper (one Python function) — “Install path A”.
A voice assistant has two halves: a mouth and ears. Since 24 October 2026 there are ears that are fast on an ordinary CPU: LT_ausys. It is Kristijonas Jakubsonas's model kmynas (NVIDIA Parakeet, trained on LIEPA-3), converted by us to ONNX and compressed so that it runs without a graphics card. Model on Hugging Face, licence CC BY 4.0. Together with this voice it makes a complete Lithuanian voice assistant for Home Assistant that runs entirely at home.
Recognises a voice command in 0.21 s. Whisper-family models (Paprika included) take 3.55 s on the same mini PC: they encode a 30-second window for every phrase, however short.
4 CPU cores and 4 GB of RAM are enough — a mini PC or a Proxmox container. The model is 0.9 GB.
On read speech (FLEURS) it gets 10.6 % of words wrong, Paprika 12.0 %. Home Assistant's default Whisper base in Lithuanian: 88 %.
Everything runs on your own machine; nothing is sent anywhere. Home Assistant connects the ears the same Wyoming way as the voice.
Paprika is a Whisper model: it runs on faster-whisper, the same engine as the usual Home Assistant Whisper add-on. LT_ausys is a different kind of model (Parakeet, ONNX), and that add-on does not offer it by itself — our proposal to add it is still waiting for review. So the ears run in their own small Wyoming server (Python environment or Docker, about 5 minutes, recipe) and are added to Home Assistant as a separate Wyoming service. Paprika can stay next to it as a backup.
Our speaker, Alisa, has been listening with them since 24 October 2026. First day: 13 commands out of 13, each in about 0.3 s, without any extra training on the owner's voice. One home's experience, not a measurement.
Before that our ears were Paprika —
also by Kristijonas Jakubsonas, a Lithuanian Whisper; it stays our choice for long dictation
(Diktuoklė). On our own test bench: generic
whisper-large-v3-turbo — 25.95 % WER, Paprika — 7.63 % (an in-domain figure: the test set comes from the same corpus).
More on Lithuanian speech recognition — setting it up in Home Assistant, the numbers and common questions — on the LT_ausys page.
Not a programmer? You don't need to understand a single command. Paste this into any AI chat (Claude, ChatGPT or another) and keep asking in your own words:
Read https://raw.githubusercontent.com/RobertasTa/reginute/main/AI_CONSULTANT_BRIEF.md
and help me install the Lithuanian voice Reginutė.
That file lists what the assistant must ask before advising and what it must not claim — so that the advice rests on facts, not guesses. Among other things it says that “the first Lithuanian speech synthesizer” would be untrue: the first was LIEPA, in 2015.
Vilnius University and the LIEPA team led by Prof. Laimutis Telksnys — for a corpus that was published rather than kept. Ten years of recordings, given away for the price of a citation.
Prof. Gražina Korvel — this voice's mother. Her group's work on VITS synthesis for Lithuanian changed this voice's path before the first training evening.
Arūnas Smaliukas — for the stress dictionary that nearly every stressed word of this voice stands on. Kristijonas Jakubsonas — for the ears. Michael Hansen and the Open Home Foundation — for a synthesizer open enough that the language of three million people could join it without asking for permission.
None of them has seen this project; no institution sponsors it. Full credits with every name and licence — at the end of the README.