lt_LT-reginute1-medium — Piper voice (Lithuanian)

Project page — an audio sample and what this is, on one page (in Lithuanian). Her younger sister Ingutė has her own page.

The first Lithuanian voice for Piper. Latvian has been in the Piper catalogue since October 2024 and Estonian since August 2026; Lithuanian never has.

• Language: Lithuanian (lt_LT)
• Voice: female — trained on the LIEPA corpus of Vilnius University, one professional actress (studio recordings, 5 121 utterances, ~3 h)
• Sample rate: 22 050 Hz · Quality: medium · Speakers: 1
• Files: lt_LT-reginute1-medium.onnx + lt_LT-reginute1-medium.onnx.json (must stay together) + phonemize_lithuanian.py with its three data tables lt_kirciai.tsv, lt_raides.tsv and lt_kreipiniai.tsv (see below — the voice does not speak without them) + skaiciu_pletiklis.py with zodziai_trumpi.txt (text normalizer) + synth_reginute.py (the recommended wrapper: sentence splitting, pauses, normalize_audio=False); SHA256SUMS covers all of them

⚠️ This voice needs its phonemizer

Lithuanian has three phonemic pitch accents, and espeak-ng places Lithuanian stress on the wrong syllable in roughly half of the words when checked against the corpus' own gold annotation. So this voice is trained with phoneme_type: text on IPA plus three accent marks (ˈ ˌ ˋ), and the text must be phonemized by phonemize_lithuanian.py together with the stress dictionary lt_kirciai.tsv (189 262 word forms, built from the LIEPA annotations and Arūnas Smaliukas' g2p-lt-lexicon). Two small tables sit beside it: lt_raides.tsv for letter names espeak-ng gets wrong when a word is spelled out (it calls both z and ž ʑˈee, so "Z." and "Ž." were read the same), and lt_kreipiniai.tsv for the vocative - Lithuanian spells "mama" the same whether you speak about her or to her, but accents it differently, and a dictionary holding one entry per spelling cannot express that.

Plain Piper with a text voice feeds the model raw letters, not IPA, and will not produce Lithuanian. A PhonemeType.LITHUANIAN contribution to piper1-gpl (in the shape of phonemize_japanese) is being prepared; until it lands, ship those files next to the model.

The phonemizer also normalizes text before phonemizing — numbers with the right case endings, clock times, units and letter-by-letter abbreviations (skaiciu_pletiklis.py, loaded automatically when it sits beside the module; pass expand_text=None to switch it off). This matters more than it sounds: left to itself, espeak-ng reads 5000 eurų as "penki tūkstantčei" and 15:00 as "tūkstantis penkišimtai:". espeak's own phonemizer normalizes text the same way, so this is the usual Piper arrangement, not an extra layer.

Recommended: normalize_audio=False — Piper's default (True) clips this voice.

Licence and lineage

• Voice files: CC-BY-4.0, derived from the meldynamics/liepa-tts corpus (CC-BY-4.0), published by Vilnius University.
• Lineage, stated openly: fine-tuned from the Piper catalogue checkpoint ru_RU-irina-medium, which is itself fine-tuned from en_US-lessac-medium. The Lessac base has an unresolved licensing question (see piper-voices discussion #94); the training data here is CC-BY-4.0, and re-basing on the LibriTTS-R (CC-BY) base model is planned.
• Code (phonemize_lithuanian.py and the tooling around it): GPL-3.0, the same licence as piper1-gpl.

Attribution (required by CC-BY-4.0)

• LIEPA corpus (2013–2015) — carried out by Vilnius University (Institute of Mathematics and Informatics; Faculty of Philology), with the Institute of the Lithuanian Language, the Lithuanian University of Educational Sciences (since 2019 the Education Academy of Vytautas Magnus University) and Šiauliai University (since 2021 the Šiauliai Academy of Vilnius University). Project lead prof. Laimutis Telksnys; corpus work and documentation Gediminas Navickas. The speaker — a professional actress — is deliberately not named here: the licence credits the corpus publishers, not the person, and we do not hold her own consent for this use. Everything anyone hears is still her. Published to Hugging Face as meldynamics/liepa-tts by MEL DYNAMICS, MB. (LIEPA-3, led by prof. Gražina Korvel, is a different and much larger corpus — it is what the ASR model below is trained on, not this voice.)
• Stress dictionary — derived from svogunas/g2p-lt-lexicon by Arūnas Smaliukas (CC BY 4.0), and from the LIEPA annotations.
• Ears, not ours but indispensable — kristijonas/paprika-whisper-lt-v3 by Kristijonas Jakubsonas (github.com/kristijonasatpro/paprika): Lithuanian ASR fine-tuned on LIEPA-3. Measured on our bench, it cut word error rate from 25.95 % (generic whisper-large-v3-turbo) to 7.63 %. This voice is the mouth of a Lithuanian assistant; his model is the ears.
• Piper — Michael Hansen and the Open Home Foundation.

Known limitations

Listed because you will meet them anyway, and finding them yourself after being told everything is fine is worse than reading them here.

None of these block ordinary use: the voice reads news, answers a smart speaker and speaks multi-minute articles without a break. They are listed so you know what you are getting — the same courtesy the author of the ASR model above extends about his own numbers.

For catalogue maintainers (piper-voices / piper1-gpl)

Nothing here is a new convention — each point follows something already merged into Piper.

1. Order: the piper1-gpl PR adding PhonemeType.LITHUANIAN first, then the piper-voices PR. The catalogue .onnx.json says "phoneme_type": "lithuanian"; a Piper without the phonemizer refuses to load the voice with a clear error instead of feeding it raw letters. The same pattern as he_IL-saspeech-medium (hebrew, piper1-gpl #244), zh_CN-chaowen-medium (pinyin, #271) and ja_JA-hi_fi_captain-medium (japanese, #274): the phoneme type names a phonemizer that lives in Piper, and the catalogue folder holds only the model and the config.
2. The phoneme id map is Piper's default with one symbol appended (ˋ = 166); ˈ and ˌ are the default 120 and 121, PAD/BOS/EOS unchanged.
3. The stress dictionary (lt_kirciai.tsv, 3 MB) is resolved through --data-dir (or next to the model) and is deliberately not in the catalogue folder — no catalogue voice ships extra files. Two merged precedents for delivering such data: bundled in the wheel via package_data like the Hebrew models (piper/hebrew/nakdimon.onnx, 20 MB, #244) — the option proposed in the piper1-gpl PR — or an archive in rhasspy/piper-checkpoints/<lang>/_resources/ downloaded on first use like g2pW (#271/#269). Until one of them is in place, take the TSV from this repo.
4. samples/speaker_0.mp3 was generated the way piper-samples/_script/generate-samples.sh does it (first line of test_sentences/lt.txt, sent to piper-samples separately), through the phonemizer. Regenerating it with a Piper that lacks the phonemizer fails; with a text config it produces noise — please keep the shipped one.

Status

In the official Piper catalogue since 2026-09-17 (PR #103); the phonemizer was merged into piper1-gpl on 2026-09-15 (PR #296) and ships with Piper 1.9.0 — from that release plain piper speaks with this voice. Until then, use the files and the wrapper in this repo.

Used in

• SpeakDream — offline voice interpreter, meeting secretary and language practice for Windows; Reginutė is its Lithuanian voice.
• A Home Assistant voice assistant (Wyoming TTS), in daily use.