LIEPA-3 — Lithuanian Speech Corpus

Didysis lietuvių kalbos garsynas (LIEPA-3)

Dataset Summary

LIEPA-3 is a large, open corpus of Lithuanian speech (~10,000 hours, ~7.5 million audio files) built for automatic speech recognition (ASR), text-to-speech (TTS) and linguistic research. It spans read, spontaneous, phonetically-annotated and dialectal speech recorded under a wide range of conditions (studio, dictaphone, radio, TV, telephone, audiobooks).

Official repository: the authoritative release is published in the CLARIN-LT repository — https://clarin-repo.lt/items/0a04648e-ba4e-4e6f-9161-a2d01e8e9fdb/full. This 🤗 dataset is a repackaging of that corpus for the datasets library.

Audio is stored once at its source quality — 44.1 kHz / 16-bit mono FLAC (lossless). Resample on the fly for your task (see below); there is no quality loss versus the source when downsampling to 16 kHz (ASR) or 24 kHz (TTS).

Supported Tasks

• automatic-speech-recognition — transcribe Lithuanian speech (audio + text). Downsample to 16 kHz at load time.
• text-to-speech — single/multi-speaker Lithuanian TTS. Filter by speaker_id, gender, source (e.g. Studio); resample to 22.05/24 kHz.

from datasets import load_dataset, Audio

ds = load_dataset("meldynamics/liepa-3", "dial", split="train")   # native 44.1 kHz

# ASR: 16 kHz on the fly (lossless downsample)
ds_asr = ds.cast_column("audio", Audio(sampling_rate=16000))

# TTS: 24 kHz
ds_tts = ds.cast_column("audio", Audio(sampling_rate=24000))

# stream without downloading everything:
ds = load_dataset("meldynamics/liepa-3", "dial", split="train", streaming=True)

Languages

Monolingual Lithuanian (lt). Transcripts are normalized to lower case without punctuation, except the dial part, which uses an extended phonetic/dialectal notation (stress ˈ ˌ, palatalization ', dialect vowels æ ə ɜ ē ɘ ɨ ā, ...).

Dataset Structure

Configurations (corpus parts)

Each of the four corpus parts is a config; load one with the name argument.

• dial — Dialectal speech — regional varieties of Lithuanian. Transcripts use an extended phonetic/dialectal notation (stress marks, palatalization, dialect-specific vowels). One phrase per file.
• phon — Phonetically annotated speech. Full recordings are sliced into phrases using the Praat TextGrid ortho (orthographic) tier; the grids' phono (phonetic-transcription) tier is aligned to each phrase by time overlap and exposed in the phono column alongside text. The source grids additionally carry phones, syll and words tiers.
• read — Read speech (prompted reading). One phrase per file with normalized, lower-cased transcripts.
• spon — Spontaneous speech (radio, TV, dictaphone, etc.). One phrase per file with normalized, lower-cased transcripts.

Data Instances

A representative example (the audio.array waveform is shown truncated):

{
  'audio': {
    'path': 'audio/dial/A/F3/A0401-SDF3/A0401-SDF3-0001.flac',
    'array': array([0.00012, -0.00008, ..., 0.00031], dtype=float32),
    'sampling_rate': 44100,
  },
  'text': "nuˈsteps ˈman' baˈtarke",
  'category': 'dial',
  'speaker_id': 'A04', 'gender': 'Female', 'age_group': '18-60',
  'lang_type': 'Spontaneous', 'source': 'Dictaphone',
  'region': 'Aukštaitija', 'county': '',
  'file_path': 'audio/dial/A/F3/A0401-SDF3/A0401-SDF3-0001.flac',
  'file_hash': '', 'segment_id': '', 'start_ms': 0, 'end_ms': 0,
  'duration_ms': 1492,
}

Data Fields

Field	Type	Description
audio	Audio	Decoded to a waveform array + sampling_rate (44100), backed by mono FLAC bytes.
text	string	Phrase transcript (normalized; dial = phonetic/dialectal).
phono	string	Phonetic transcription of the phrase (phon only; SAMPA-style, from the TextGrid phono tier). Empty for other configs.
category	string	Corpus part: dial / phon / read / spon.
speaker_id	string	Speaker code, where derivable from the path.
gender	string	Male / Female.
age_group	string	0-12 / 13-17 / 18-60 / 60+.
lang_type	string	Read / Spontaneous / Mixed.
source	string	Studio / Dictaphone / Radio / TV / Phone / Audiobook / Unknown.
region	string	Dialect region (dial): Aukštaitija / Dzūkija / Suvalkija / Žemaitija.
county	string	County of the speaker (read).
file_path	string	Original path of the clip inside the corpus archive.
file_hash	string	Content-hash id of the source recording (phon / spon).
segment_id	string	Phrase index within the source recording (phon).
start_ms / end_ms	int32	Phrase boundaries within the source recording (phon).
duration_ms	int32	Clip duration in milliseconds.

Data Splits

Each config has a single train split. Exact per-config row/byte counts appear in the viewer's Dataset Structure panel. Full-corpus totals:

Part	Duration (h:m:s)	Speakers	Phrases	Words	Files	Size (FLAC 44.1k)
Dialectal (dial)	100:29:24	117	111,970	916,343	111,970	15 GB
Phonetic (phon)	501:08:59	500+	1,084,228	4,383,938	145,644	74 GB
Read (read)	5015:37:12	7,125	2,699,385	26,984,811	2,699,385	578 GB
Spontaneous (spon)	4937:25:27	5,000+	4,527,544	43,135,179	4,527,544	608 GB

Audio Format

FLAC, 44.1 kHz, 16-bit, mono (source quality). Phrase durations (source): dial avg 3.23 s, read avg 6.69 s, spon avg 3.93 s.

Dataset Creation

Curation rationale. Advance Lithuanian speech technology (ASR/TTS) and provide a phonetically representative, openly-licensed resource for a comparatively low-resource language.

Source data. Recordings gathered from many sources/conditions — studio, dictaphone, radio and TV, telephone and audiobooks — to maximize acoustic diversity. Encoding codes in file paths capture language type, source, gender, age group and region/county.

Annotations. Phrase-level transcription covers all 10,000 h. The phonetic (phon) part (500 h) additionally provides word, syllable and phoneme annotations as Praat TextGrid files (phoneme set per Kasparaitis 2005, 92 units). The dial part adds dialectal phonetic transcription.

Personal and sensitive information. Speakers are referenced only by anonymized codes. Spontaneous/broadcast material may mention real people or events; use responsibly.

Considerations for Using the Data

Social impact. Openly-licensed, large-scale Lithuanian speech data — a comparatively low-resource language — lowers the barrier to accessible voice interfaces, transcription and assistive technology, and supports Baltic-language speech research.

Distribution & biases. Aggregate gender balance by part (hours):

Part	Total h	Male h (%)	Female h (%)
dial	100	23 (23%)	77 (77%)
phon	501	291 (58%)	210 (42%)
read	5,015	2,154 (43%)	2,842 (57%)
spon	4,937	2,552 (52%)	2,385 (48%)

Source/condition coverage is uneven (read speech dominates total hours); account for domain and speaker imbalance when training/evaluating.

Known limitations. phon clips are sliced on the TextGrid ortho tier; normalized transcripts drop casing/punctuation (except dial); demographics are only as complete as the original path encoding.

Additional Information

Official Release

The authoritative dataset is distributed via the CLARIN-LT repository: https://clarin-repo.lt/items/0a04648e-ba4e-4e6f-9161-a2d01e8e9fdb/full

Original Documentation

The authoritative corpus documentation is bundled in this repository:

• docs/LIEPA-3-README.md — full description: labeling codes, per-part directory structure, annotation formats (Lithuanian).
• docs/LIEPA-3-README.pdf — PDF version.
• LICENSE — full CC BY 4.0 legal text.

Dataset Curators

Created 2025–2026 by Vytautas Magnus University (VDU), Vilnius University (VU) and the Institute of the Lithuanian Language (LKI).

Licensing Information

Released under Creative Commons Attribution 4.0 International (CC BY 4.0). Share and adapt for any purpose with appropriate credit to the LIEPA-3 authors.

Citation

@misc{liepa3,
  title        = {Didysis lietuvių kalbos garsynas (LIEPA-3)},
  author       = {Vytautas Magnus University and Vilnius University and
                  Institute of the Lithuanian Language},
  year         = {2026},
  note         = {Large Lithuanian speech corpus, ~10,000 hours},
  howpublished = {CC BY 4.0}
}

Contributions

Packaged for the 🤗 datasets library from the original FLAC corpus. See the bundled original documentation for authoritative details.