Datasets:

streams-sro
/

trepal

license: cc-by-4.0
language:
  - sk
task_categories:
  - automatic-speech-recognition
pretty_name: 'TrePal: Slovak parliament speech 2024–2026'
size_categories:
  - 100K<n<1M
tags:
  - speech
  - slovak
  - parliament
  - asr
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*

TrePal

Talking Representatives: Elected Parliament, Aligned & Labelled.

The sequel to SloPal: first they drank, now they talk.

TrePal has 471.5 hours of Slovak speech from the plenary sessions of the National Council of the Slovak Republic (NR SR), from November 2024 to September 2026. That is the period right after SloPal ends.

140,420 clips of 1.5–25.4 s.
Each clip is aligned to the official, human-checked stenographic record.
Each clip comes with its speaker, MP ID, parliamentary club and session.

Released by Streams s.r.o., the makers of Gaston.

Quick start

from datasets import load_dataset

ds = load_dataset("streams-sro/trepal", split="test")
print(ds[0]["text"], ds[0]["speaker"], ds[0]["club"])
ds[0]["audio"]  # 16 kHz mono

At a glance

Split	Recordings	Clips	Hours	Speakers
train	193	136,708	458.8	190
validation	2	1,771	5.93	63
test	2	1,941	6.77	73
total	197	140,420	471.5	190

Audio is FLAC, 16 kHz mono. Clips are 1.5–25.4 s (mean ≈ 12 s), cut at word boundaries found by forced alignment.
The splits are by whole recordings (one sitting = one recording). Test and validation are each 2 recordings, and nothing in them is in train.
- The same MPs speak in every split, so the test measures new speeches by familiar speakers, not unseen speakers.
Sessions 23–61 of the 9th electoral term, 2024-11-26 → 2026-09-17.
- Sessions 55–60 are missing: their records weren't human-checked yet when we collected the data.
- We may add them in a later version.

Fields

Field	Description
`id`	`<recording>_<speech item>_<segment>`
`audio`	FLAC 16 kHz mono
`text`	The training text: the official record's words, with single-meaning abbreviations written out as spoken ("napr." → "napríklad", "resp." → "respektíve", "č." → "číslo"). Numbers stay as the record writes them.
`text_record`	The same words exactly as in the official record. They differ from `text` in 1.8 % of clips.
`duration`, `start`, `end`	Seconds; `start` / `end` are positions in the full sitting recording
`recording`	The sitting recording the clip comes from (date and time of the NR SR video)
`session`, `date`	Session number and ISO date
`speech_type`	The NR SR item type, e.g. "Vystúpenie v rozprave" (debate speech), "Vystúpenie s faktickou poznámkou" (short reply), "Vstup predsedajúceho" (chair)
`speaker`	Speaker name as in the record
`poslanec_id`	The MP's ID on nrsr.sk. Empty for non-MPs (ministers, guests).
`club`, `party_list`, `region`	The MP's parliamentary club, election list and region, from the nrsr.sk profile as of October 2026. MPs who changed club show their current one.

We publish only metadata about the public role. Birth dates, residence and nationality are left out on purpose.

How it was made

Source. Audio of all plenary sittings from the NR SR media portal (tv.nrsr.sk), 319 recordings, 1,203 h. Transcripts per speech came from the same portal. Only speeches marked "Skontrolovaný text" (checked text) are used. Speeches still carrying the portal's unchecked AI transcript (22 h after alignment) were dropped.
Alignment. The portal's speech timestamps sit on a 15-minute grid, so they can't be used. We align each whole sitting as one word stream:
- anchors are word 3-grams that are unique both in the record and in the greedy transcript of MMS-1b-all, kept in order (longest increasing subsequence);
- CTC forced alignment runs between the anchors, with free "star" tokens at speaker changes and at stage notes.
- This handles hour-long recesses and the chair's repeated formulas, which break a simple left-to-right aligner.
Clips. Speech is cut into clips of up to ~25 s at word boundaries. A clip never starts or ends on a number.
- Dropped: clips shorter than 1.5 s, clips with a low mean alignment score, and clips with a speech rate outside 6–25 characters per second.
- Chair lines and vote results ("143 prítomných, 76 za…") are typed into the record but never spoken, so they fall out on their own.
Text.
- Bracketed stage notes ("(Potlesk.)", "(Prestávka.)") and inline speaker labels are removed.
- Abbreviations with a single spoken form are expanded.
- Clips with case-dependent abbreviations ("§", "ods.", "Z. z.", "mil.", "tzv."…) are left out, because their spoken form can't be recovered without guessing.
- Unicode is NFC-normalized.
Independent check. An independent ASR model (kinit/whisper-large-v3-sk) transcribed all kept clips and wasn't used for selection. CER between our text and its transcript: median 0.3 %, 90 % of clips ≤ 7.2 %. A manual look at the outliers found mostly that model's own errors.

Of the 1,203 h of sitting audio, 485.7 h became aligned clips from checked speeches. The text rules in step 4 leave 471.5 h.

Who speaks

Hours by parliamentary club (club as of October 2026, all splits):

Club	Hours	Share
Klub Progresívne Slovensko	144.5	31 %
Klub SLOVENSKO - ZA ĽUDÍ	89.2	19 %
Klub Sloboda a Solidarita	80.3	17 %
(no club / not an MP)	66.1	14 %
Klub KDH	34.7	7 %
Klub SMER - sociálna demokracia	25.8	5 %
Klub HLAS - sociálna demokracia	23.9	5 %
Klub Slovenská národná strana	7.0	1 %

Three quarters of the speaking time comes from the opposition clubs (PS, SLOVENSKO, SaS, KDH) and about a tenth from the governing parties' clubs (SMER, HLAS, SNS).

The "no club" row is mostly government ministers, plus MPs who belong to no club.
The table shows how the floor time was used in 2024–26, not how big the clubs are.
If you need balance across speakers or parties, sample by speaker or club.

Benchmark

WER in %, lower is better, on the full TrePal test split (1,941 clips, 6.77 h).

Decoding: CTranslate2 fp16, beam 5, language sk, no prompt.
Text normalization: lowercase, punctuation removed, numbers converted to digits on both sides.

Model	WER	CER
openai/whisper-large-v3	8.42	3.01
kinit/whisper-large-v3-sk	2.41	1.29
Gaston Slovak model (gaston-sk-v2)¹	1.47	0.62

¹ Available through Gaston and its API; the weights aren't public. It was trained on the TrePal train split (among other data) and has never seen the test recordings.

Limitations

The record isn't verbatim. The stenographers smooth the speech: they drop fillers, false starts and repetitions, and lightly correct grammar.
- So the text is clean written Slovak of what was said, not a phonetic transcript.
- Some clips contain words the record leaves out. The aligner absorbs short ones.
Formal register only. These are parliamentary speeches, read or semi-spontaneous, by about 190 speakers. Don't expect it to cover casual or regional speech.
Speaker labels follow the record. Interjections from the floor are usually not transcribed, and the audio of a clip may contain them.
Seven source recordings are truncated at the source. Speeches past the cut-off are simply missing.
club / party_list describe the MP in October 2026, not on the day of the speech.

License and attribution

CC BY 4.0. The audio and transcripts come from the National Council of the Slovak Republic (Národná rada Slovenskej republiky, nrsr.sk, tv.nrsr.sk).

The segmentation, alignment, cleaning and metadata are by Streams s.r.o.
If you use TrePal, please credit both the NR SR and this dataset.

Citation

@misc{trepal2026,
  title        = {TrePal: Slovak Parliament Speech 2024--2026, Aligned and Labelled},
  author       = {{Streams s.r.o.}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/streams-sro/trepal}}
}

Thanks to the authors of SloPal (NaiveNeuron) for showing the way, and to the NR SR for publishing its sittings and records openly.