Digisensus/lithuanian-dialect-speech-liepa-3-100h-punctuated

Lithuanian Dialect Speech 100 h: punctuated, cased, numbers as digits (written form)

Spontaneous Lithuanian dialect speech from all four regions, with three transcripts per clip: written form (punctuation, capitalisation, numbers as digits), normalised, and the original phonetic transcription with stress marks. Dialect word forms are kept as spoken in every layer.

Released by [Digisensus](https://digisensus.com), a conversational intelligence company from Vilnius, Lithuania, building custom enterprise ASR and LLM solutions for call transcription and conversation analytics.

Companion dataset: Lithuanian Phone Speech 429 h (same conventions, compatible columns).

Summary

• —100.5 hours, 111,970 utterances, 117 speakers of spontaneous Lithuanian dialect speech from Aukštaitija, Žemaitija, Dzūkija and Suvalkija: the complete dialect part of the openly licensed LIEPA-3 corpus.
• —Written-form transcripts for every clip, so a speech recogniser can learn to output readable text directly, with no separate punctuation, casing or inverse text normalisation (ITN) step.
• —Three layers side by side: written form, a normalised lowercase transcript, and the original LIEPA-3 phonetic transcription (stress, palatalisation, dialect vowel qualities).
• —Speaker IDs and regions for every clip, so you can build speaker-disjoint or per-region splits.
• —Ready to use: Parquet with embedded 16 kHz audio, CC BY 4.0.

Quick start

from datasets import load_dataset

ds = load_dataset("Digisensus/lithuanian-dialect-speech-liepa-3-100h-punctuated", split="train")
sample = ds[0]
print(sample["text"])             # written form
print(sample["text_normalized"])  # lowercase, no punctuation, numbers as words
print(sample["text_phonetic"])    # original LIEPA-3 phonetic transcription
audio = sample["audio"]           # {"array": ..., "sampling_rate": 16000}

Decoding audio needs datasets>=4 with torchcodec installed. To read the raw FLAC bytes without decoding: from datasets import Audio, then ds.cast_column("audio", Audio(decode=False)). Streaming: load_dataset("Digisensus/lithuanian-dialect-speech-liepa-3-100h-punctuated", split="train", streaming=True).

Dataset structure

Example:

{
  "id": "S1301-SDF3-0004",
  "text": "Per 3 klases buvu 10 mokinių.",
  "text_normalized": "per tris klases buvu dešim mokinių",
  "text_phonetic": "per ˈtris klaˈses ˈbuvu ˈdæšim mokiˈnių",
  "duration": 2.19,
  "speaker_id": "S13",
  "region": "Suvalkija",
  "gender": "female",
  "age_group": "18-60",
  "liepa3_path": "audio/dial/S/F3/S1301-SDF3/S1301-SDF3-0004.flac"
}

Rows are sorted by id, so consecutive rows are consecutive utterances of one recording session. The data is stored as Parquet shards of about 500 MB in data/.

The columns of the phone dataset are a subset of these: drop text_phonetic, speaker_id and region to concatenate the two. Note that text_normalized there is LIEPA-3's own transcript; here it is derived from text_phonetic.

Splits: bring your own

The dataset is published as a single `train` split, with no validation or test part. Choose held-out data to suit your task. To measure how a model handles unseen speakers, hold out whole speakers, not random utterances; a speaker's clips share voice, subdialect and topic.

held_out = {"A07", "D05", "S02", "Z13"}   # e.g. one or more speakers per region
test = ds.filter(lambda s: s in held_out, input_columns="speaker_id")
train = ds.filter(lambda s: s not in held_out, input_columns="speaker_id")

speaker_id is stable across sessions: three speakers were recorded in more than one session (A04: 12 sessions, Z24: 4, D04: 2).

Regions and speakers

region is the LIEPA-3 region label. LIEPA-3 labels regions, not subdialects. Linguistically, Dzūkija and Suvalkija speech belongs to the Aukštaitian dialect (southern and western Aukštaitian); Žemaitija is Samogitian.

Utterances are short: 0.7–30 s, 3.2 s on average, 916,000 words in total.

How it was built

1. 1.Selection. All 111,970 clips of LIEPA-3's dialect (dial) part are included; nothing was filtered out.
2. 2.Audio. Resampled from the original 44.1 kHz to 16 kHz, 16-bit mono FLAC. The content is not cut, trimmed or otherwise edited.
3. 3.Normalised text. text_normalized is derived from text_phonetic by a fixed rule set: stress and palatalisation marks removed, dialect vowel letters mapped to Lithuanian letters (see below).
4. 4.Conversion. Each normalised transcript was converted to written form following the transcription conventions: numbers by meaning, capitalisation, punctuation. The conversion was semi-automatic, using a locally run Gemma model.
5. 5.Checks. Every converted transcript was compared automatically with its normalised original: the sequence of spoken words must be the same, and every figure must be derivable from the numerals actually spoken. Failed conversions were retried. About 1,400 that still failed were reviewed by an operator, and 1,370 of those were accepted with deviations (see Known limitations).
6. 6.Targeted review. Detectors flagged 2,796 utterances with conversions that are easy to get wrong in dialect speech: numeral idioms written as digits, the pronoun "vienas" written as "1", dialect pronouns capitalised as names, clock times, unit symbols, ordinal endings, and the preposition "no" (= "nuo"). An LLM reviewer (Claude Sonnet) checked each one against the conventions below. Its 1,129 corrections that passed the word and number checks were applied; corrections that dropped or added words, removed an acronym or introduced an unverifiable number were discarded.

Phonetic notation and the normalised text

text_phonetic is the LIEPA-3 transcription, unchanged. Besides Lithuanian letters it uses:

The mapping is approximate by design: it gives a plain-letter target for ASR and loses vowel quality, stress and softness. Use text_phonetic when those matter.

Transcription conventions

text keeps the spoken words of the transcript, in the same order and in their dialect form ("kap", "tep", "buva", "dzienas", "anėi"). Only three things differ from text_normalized:

1. 1.spelled-out numbers are written as digits or symbols, by their meaning in context;
2. 2.capitalisation;
3. 3.punctuation.

The rules are the same as in the phone dataset. Dialect numerals are converted by their value, and ordinal endings follow the spoken form:

Numbers by meaning

When several readings compete, the priority is: official name > identifier/code > date > clock time > currency > percentage > measurement/duration > ordinal > cardinal. In ambiguous contexts the form that adds the least interpretation is used.

Approximation words stay as words ("apie 5", "kokie 3"), and so do vague quantities with no figure ("šimtus", "keli", "pora").

Casing and punctuation

• —Every transcript starts with a capital letter, including fragments that begin mid-sentence.
• —Proper names are capitalised; month names stay lowercase, as in Lithuanian.
• —Sentences end with ., ? or !; fragments also end with a full stop (all but 44 utterances end with a punctuation mark). Commas follow Lithuanian writing rules.

Figures in the data

8.2 % of utterances contain at least one digit. Share of utterances containing: ordinals 1.6 %, unit symbols 0.8 %, ranges 0.2 %, clock times 0.2 %, years with "m." 0.15 %, decimals 0.13 %, grouped thousands 0.11 %, dates with "d." 0.10 %, currency 0.04 %, percentages 0.02 %.

Known limitations

• —No validation or test split (see Splits).
• —Word-level differences from the normalised transcript: 99.2 % of utterances pass the automatic word and number checks against text_normalized. The remaining 0.8 % (869 utterances) were accepted in review with deviations. Most are dialect unit words written as symbols ("mėtru" → "m", "auru" → "€") or dialect spellings the checker does not recognise, but a few may contain real errors in a number or unit (e.g. "valanda" → "1 val.", where no number was spoken). Use text_normalized if you need the exact spoken word sequence.
• —Less consistent than the phone dataset: the conversion model was a smaller, locally run one, and the targeted review covered only flagged utterances. Remaining inconsistencies:
• —dialect pronouns capitalised mid-sentence as if they were names ("Ana", "Anas", "Anėi" for "ji", "jis", "jie") in about 100 utterances (some are real names); about 2,200 other occurrences are lowercase;
• —the choice between digit and word for "vienas / viena" (count or pronoun) and between a clock time and a plain number is a judgement call, and similar utterances may differ;
• —ordinal endings follow the spoken dialect form, so the same ordinal can appear with different endings ("1-os", "1-uos"; "1-as", "1-s").
• —Normalised text is approximate: the vowel mapping in text_normalized and text merges distinct dialect sounds (e.g. æ, ɜ, ɘ all become e).
• —Final punctuation of fragments: unlike the phone dataset, where about 15 % of utterances have no final punctuation mark, fragments here end with a full stop.
• —Audio: dictaphone recordings made under varying conditions, downsampled from 44.1 kHz. The full-band originals are in LIEPA-3 (see liepa3_path).
• —Demographics: 77 % of the hours are female speech, and 44 % are women over 60. There are no speakers under 18. Male speech is scarce for Dzūkija (3.5 h) and Suvalkija (3.8 h).
• —Uneven speakers: from under a minute to about 2 hours per speaker (median 0.85 h).

Found an error? Open a discussion in the Community tab. Corrections go into the next version.

Source and attribution

The audio, the phonetic transcripts and the speaker metadata come from LIEPA-3 — Didysis lietuvių kalbos garsynas (the large Lithuanian speech corpus), created in 2025–2026 by Vytautas Magnus University, Vilnius University and the Institute of the Lithuanian Language, released under the Creative Commons Attribution 4.0 International licence. Original release: CLARIN-LT repository.

This dataset is not affiliated with, sponsored or endorsed by the LIEPA-3 authors or their institutions.

Changes made to the original material

• —Selection: the complete dialect part of LIEPA-3 (dial: 111,970 clips, 117 speakers). Other corpus parts (read, spontaneous, phonetic) are not included.
• —Audio: resampled from 44.1 kHz to mono 16 kHz FLAC. The content is not cut, trimmed or otherwise edited.
• —Transcripts: the original phonetic transcript is kept unchanged in text_phonetic. A normalised transcript (text_normalized) and a written-form transcript (text) were added.
• —Metadata: speaker, region, gender and age group were decoded from the LIEPA-3 file and folder names.

About Digisensus

Digisensus is a conversational intelligence company founded in 2025 in Vilnius, Lithuania (EU). We build custom enterprise ASR and LLM solutions for conversations across calls, email, chat and messaging: accurate phone-call transcription, customer journey analytics, and conversation intelligence extraction at scale (thousands of conversations in a single analysis).

Accurate transcription of real conversations in smaller languages depends on data that did not exist in the open. Regional speech is part of that: many callers do not speak the standard language. We built this dataset for our own Lithuanian speech recognition work and release it so that researchers and engineers working on Lithuanian and other low-resource languages can build on it.

Licence and required attribution

This dataset is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). The full licence text is in `LICENSE`.

The written-form and normalised transcripts, the curation and the dataset compilation are © 2026 Digisensus, licensed under CC BY 4.0. The audio, phonetic transcripts and metadata are from LIEPA-3, also CC BY 4.0.

CC BY 4.0 requires attribution. When you use, redistribute or adapt this dataset, including in trained models, papers, products and derived datasets, credit both sources as follows:

Lithuanian Dialect Speech 100 h (written-form transcripts) by Digisensus, licensed under CC BY 4.0. Derived from LIEPA-3 by Vytautas Magnus University, Vilnius University and the Institute of the Lithuanian Language, licensed under CC BY 4.0.

In a model card, dataset card or paper, a citation (below) together with a link to digisensus.com satisfies this. As the licence requires, link to the licence and say if you changed the data.

Citation

@misc{digisensus_lithuanian_dialect_speech_2026,
  title        = {Lithuanian Dialect Speech 100 h: punctuated, cased, numbers as digits (written form)},
  author       = {{Digisensus}},
  year         = {2026},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/datasets/Digisensus/lithuanian-dialect-speech-liepa-3-100h-punctuated},
  note         = {Digisensus, https://digisensus.com. Derived from LIEPA-3 (Vytautas Magnus
                  University, Vilnius University, Institute of the Lithuanian Language), CC BY 4.0}
}