Digisensus/lithuanian-phone-speech-liepa-3-429h-punctuated

Lithuanian Phone Speech 429 h: punctuated, cased, numbers as digits (written form)

Transcripts are in written form, not normalised: punctuation, capitalisation, and numbers, dates, times and amounts as digits ("2026 m. rugsėjo 6 d., 9:30", "65 000 €", "12,5 %"). The original normalised text is included too.

text	text_normalized
Varšuva 85 % sugriauta.	varšuva aštuoniasdešim penki procentai sugriauta
Keliais eurais arba 10 € daugiau kaip valytojos.	keliais eurais arba dešimčia eurų daugiau kaip valytojos

Released by Digisensus, a conversational intelligence company from Vilnius, Lithuania, building custom enterprise ASR and LLM solutions for call transcription and conversation analytics.

Companion dataset: Lithuanian Dialect Speech 100 h (spontaneous dialect speech from all four regions, same conventions, compatible columns).

Summary

• 429 hours, 417,000 utterances of spontaneous Lithuanian telephone speech, from 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.
• Both forms side by side: the original normalised transcript is kept, so the data also works for normalised-output ASR and as a parallel corpus for Lithuanian text normalisation and ITN.
• Documented conventions: numbers are written by their meaning in context (date, clock time, currency, measurement, ordinal, code and so on), following the rules described below.
• Verified: every transcript was checked automatically against the spoken words and numerals; transcripts that failed a check were reviewed.
• Ready to use: Parquet with embedded 16 kHz audio, train / validation / test splits, CC BY 4.0.

To our knowledge, this is the largest openly licensed Lithuanian telephone speech dataset with written-form transcripts.

Quick start

from datasets import load_dataset

ds = load_dataset("Digisensus/lithuanian-phone-speech-liepa-3-429h-punctuated")
sample = ds["train"][0]
print(sample["text"])             # written form
print(sample["text_normalized"])  # original LIEPA-3 transcript
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-phone-speech-liepa-3-429h-punctuated", split="train", streaming=True).

Dataset structure

Field	Type	Description
id	string	The 12-character clip ID from the LIEPA-3 file name
audio	Audio	Mono 16 kHz FLAC, embedded
text	string	Written-form transcript (see transcription conventions)
text_normalized	string	Original LIEPA-3 transcript: lowercase, no punctuation, numbers as words
duration	float32	Clip length in seconds
gender	string	female / male (from LIEPA-3 metadata)
age_group	string	0-12, 13-17, 18-60, 60+ (from LIEPA-3 metadata)
liepa3_path	string	Path of the clip in the original LIEPA-3 release

Example:

{
  "id": "0051d0b44f0b",
  "text": "Kaip atrodė Varšuva anksčiau, nes aš ją visiškai atrandu iš naujo, nes Varšuva 85 % sugriauta.",
  "text_normalized": "kaip atrodė varšuva anksčiau nes aš ją visiškai atrandu iš naujo nes varšuva aštuoniasdešim penki procentai sugriauta",
  "duration": 6.54,
  "gender": "female",
  "age_group": "18-60",
  "liepa3_path": "audio/spon/S/P/F3/00/51/0051d0b44f0b-SPF3.flac"
}

The data is stored as Parquet shards of about 500 MB in data/.

Splits

Split	Utterances	Hours
train	408,641	420.2
validation	4,153	4.3
test	4,206	4.4
total	417,000	428.9

Splits are deterministic: sha1(id) mod 100 — bucket 0 is test, bucket 1 is validation, the rest is train. LIEPA-3 does not publish speaker identifiers for this part of the corpus, so the splits are not speaker-disjoint: the same caller can appear in more than one split. Treat validation/test scores as in-domain estimates, not as speaker-independent results.

Speakers

Gender / age	Utterances	Hours
male 18–60	227,504	229.8
female 18–60	126,938	139.4
male 60+	44,227	41.8
female 60+	17,767	17.5
female 13–17	228	0.2
male 0–12	128	0.1
female 0–12	135	0.1
male 13–17	73	0.1

Utterances are short: 0.8–30 s, 3.7 s on average, 3.6 million words in total.

How it was built

1. Selection. Spontaneous telephone recordings were selected from LIEPA-3 (see changes made) and converted to mono 16 kHz FLAC.
2. Conversion. Each normalised transcript was converted to written form following the transcription conventions: numbers by meaning, capitalisation, punctuation. The conversion was semi-automatic.
3. Word check. Every converted transcript was compared automatically with its original. The sequence of spoken words must be the same; only numerals, casing and punctuation may differ.
4. Number check. Every figure in the written form (digits, grouped thousands, decimals, clock times, years, Roman numerals) was verified to be derivable from the numerals actually spoken. This catches wrong values, merged numbers and invented figures.
5. Review. Transcripts that failed a check were reviewed, and corrected where the conversion was wrong. Utterances flagged by the number check were reviewed by hand.

Transcription conventions

text keeps the spoken words of the original transcript, in the same order. Only three things differ from text_normalized:

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

By convention, words are not added, removed, reordered, corrected or paraphrased. Fillers, colloquial and non-standard forms stay as spoken ("nu", "biškį", "ane", "mhm", "jo"). For the small share of utterances that deviate, see Known limitations.

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.

Category	Spoken (normalised)	Written
Cardinals — digits, including 1–9	penki autobusai	5 autobusai
Ordinals — digit, hyphen, spoken ending	pirmoji vieta, liepos trisdešimt pirmąją	1-oji vieta, liepos 31-ąją
Large numbers — groups of three, space-separated	du šimtai penkiasdešimt tūkstančių	250 000
Decimals — decimal comma	vienas kablelis penki, trys su puse	1,5, 3,5
Signed numbers	minus penki	-5
Percentages — space before %	dvylika kablelis penki procento	12,5 %
Currency	penki eurai dvidešimt centų, šešiasdešimt penkių tūkstančių eurų	5,20 €, 65 000 €
Measurements — unit symbols	penki kilometrai, penki kvadratiniai metrai	5 km, 5 m²
Temperature	minus penki laipsniai Celsijaus	-5 °C
Clock time — 24 h H:MM, no leading zero	pusę dešimtos, devintą vakaro, be penkiolikos dvylika	9:30, 21:00, 11:45
Durations — Lithuanian abbreviations	dvi valandas, trisdešimt minučių, penkias sekundes	2 val., 30 min., 5 sek.
Dates	rugpjūčio šeštą dieną; du tūkstančiai dvidešimt šeštų metų rugsėjo šeštą	rugpjūčio 6 d.; 2026 m. rugsėjo 6 d.
Years	du tūkstančiai keturioliktais metais, devyniasdešimtųjų	2014 m., 90-ųjų
Fractions	viena antroji, pusantro litro	1/2, 1,5 l
Ranges — en dash between figures	šešias aštuonias dienas; nuo penkių iki dešimties	6–8 dienas; nuo 5 iki 10
Phone numbers, codes, PINs — digit strings, leading zeros kept; spoken groups kept	nulis nulis penki; šeši aštuoniasdešimt vienuolika	005; 6 80 11
Addresses, routes, model numbers	Gedimino prospektas dvidešimt; penktas autobusas (route)	Gedimino prospektas 20; 5 autobusas
Official names	Jonas Paulius antrasis	Jonas Paulius II
Centuries and other Roman-numeral conventions	dvidešimt pirmas amžius	XXI a.

Stays as words:

• approximation and comparison words around a figure: "apie 5", "daugiau nei 20", "kokie 3";
• vague quantities with no figure: "šimtus žmonių", "tūkstančiai", "keli", "keliolika", "pora";
• idioms and pronoun-like uses of numerals: "vienas iš", "nė vieno", "viena vertus", "pirmą kartą";
• "pusė" on its own, unless it is part of a figure.

Casing and punctuation

• Every transcript starts with a capital letter, including fragments that begin mid-sentence.
• Sentences start with a capital letter; proper names (people, places, companies, brands) are capitalised; month names stay lowercase, as in Lithuanian.
• Acronyms are written in capitals: BVP, PVM, ES, JAV.
• Sentences end with ., ? or !. Commas follow Lithuanian writing rules (subordinate clauses with "kad", "nes", conjunctions "bet", "o", enumerations, vocatives, "nu,"). Fillers are not over-punctuated.
• Ordinal endings use a plain hyphen (31-ąją), ranges use an en dash (5–10 km); Lithuanian quotation marks „…“ are used.

Figures in the data

7.9 % of utterances contain at least one digit. Share of utterances containing: ordinals 1.5 %, percentages 0.6 %, grouped thousands 0.6 %, ranges 0.5 %, years with "m." 0.5 %, currency 0.3 %, decimals 0.2 %, clock times 0.1 %.

Known limitations

• Not speaker-disjoint splits (see Splits).
• Word-level differences from the original transcript: 98.9 % of utterances pass the word-sequence check against text_normalized. The remaining 1.1 % (about 4,700 utterances) were accepted in review with small deviations, mostly:
  • foreign names and terms in their standard spelling instead of the phonetic one ("fon der lajon" → "Von der Leyen", "kovid devyniolika" → "COVID-19", "luganską" → "Luhanską");
  • a dropped repeated or disfluent word;
  • an occasional grammatical correction of the spoken form. Use text_normalized if you need the exact spoken word sequence.
• Garbled source words: LIEPA-3 transcripts sometimes contain misspelled or fused numerals ("aštuoniasdešitm", "ikipenkiolikos"). The written form gives the intended value ("80 %", "iki 15").
• Punctuation of fragments: about 15 % of utterances are cut mid-sentence and have no final punctuation mark.
• Remaining inconsistencies: utterances with an ambiguous reading may differ in formatting from similar ones: e.g. short spoken years ("dvidešimt pirmais" → "2021 m." vs "21-ais"), ordinal endings ("3-as" vs "3-ias"), date word order, duration vs clock time for bare hours, or whether an idiomatic "vienas" was written as a digit.
• Audio channel: telephone recordings of varying quality, stored at 16 kHz; the effective bandwidth is often narrowband.
• Demographics: adults 18–60 dominate; children and teenagers are under 0.2 % of the data; male speech is about 63 % of the hours.

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

Source and attribution

The audio and the normalised transcripts 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: only the spontaneous telephone recordings of LIEPA-3's spon part (speech type S, source P; 489,853 clips) were used. From those, a random selection of 417,000 clips (428.9 h) with a duration between 0.5 s and 30 s is included. Other speech types, sources and corpus parts (read, dialect, phonetic) are not included; the dialect part is published as Lithuanian Dialect Speech 100 h.
• Audio: converted to mono 16 kHz FLAC. The content is not cut, trimmed or otherwise edited.
• Transcripts: a new written-form transcript (text) was added for every clip, following the conventions below. The original normalised transcript is kept unchanged in text_normalized.
• Metadata: gender and age group were decoded from the LIEPA-3 file name codes; a train/validation/test split was added.

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 phone calls in smaller languages depends on data that did not exist in the open. 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.

Coming next: speech recognition models trained on this data, with evaluation results and findings, will be published on the Digisensus Hugging Face page.

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 transcripts, the curation and the dataset compilation are © 2026 Digisensus, licensed under CC BY 4.0. The audio and normalised transcripts 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 Phone Speech 429 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_phone_speech_2026,
  title        = {Lithuanian Phone Speech 429 h: punctuated, cased, numbers as digits (written form)},
  author       = {{Digisensus}},
  year         = {2026},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/datasets/Digisensus/lithuanian-phone-speech-liepa-3-429h-punctuated},
  note         = {Digisensus, https://digisensus.com. Derived from LIEPA-3 (Vytautas Magnus
                  University, Vilnius University, Institute of the Lithuanian Language), CC BY 4.0}
}