CoolFace
Datasetpublic

porupski/ParlaSpeech-HR-benchmark_v3

ParlaSpeechHR Benchmark v3 A curated benchmark dataset of 22,008 Croatian parliamentary speech clips extracted from ParlaSpeech-HR v3. Each clip includes aligned audio (WAV) and TextGrid annotations for linguistic analysis. Contents 22,008 audio segments (various durations) 17,622 clips with complete TextGrid triplets: .align (word-level boundaries via WordAlign tier) .stress (primary stress frame labels; derivative of .align) .pause (filled pause annotations… See the full description on the dataset page: https://huggingface.co/datasets/porupski/ParlaSpeech-HR-benchmark_v3.

sourceHugging Faceupdated 3mo agoView on Hugging Face
0likes244downloads
Dataset Card

ParlaSpeechHR Benchmark v3

A curated benchmark dataset of 22,008 Croatian parliamentary speech clips extracted from ParlaSpeech-HR v3. Each clip includes aligned audio (WAV) and TextGrid annotations for linguistic analysis.

Contents

  • —22,008 audio segments (various durations)
  • —17,622 clips with complete TextGrid triplets:
  • —.align (word-level boundaries via WordAlign tier)
  • —.stress (primary stress frame labels; derivative of .align)
  • —.pause (filled pause annotations with rougher word-tier alignment)
  • —4,386 clips with .pause annotations only (align & stress unavailable from source)

Data format

Two datasets configs, each with a single train split. Task splits are per-task columns, not HF splits — the same utterance can be gender/train and age/dev at the same time.

`default` — what the viewer opens; everything you need to browse, filter, and train on utterance-level tasks.

ColumnTypeNotes
instance_idstringe.g. ParlaMint-HR_2021-01-29-1.u30991_41-251
audioAudio(sampling_rate=16000)16 kHz mono, inlined bytes
text, audio_length, text_start, text_end, audio_start, audio_endscalarutterance-level anchors
speaker_id, speaker_name, speaker_gender, speaker_birth, speaker_role, speaker_mp, speaker_minister, speaker_party, party_status, party_orientationscalarspeaker & party. party_orientation is numeric (−3.0 … +3.0), mapped from the textual label
text_id, session_id, title, date, body, term, session, meeting, sitting, agenda, subcorpus, langscalarParlaMint provenance
sentiment_logit, sentiment_3, sentiment_6scalarParlaSent outputs
sentiment_regressionfloatnumeric map of sentiment_6 (−2.5 … +2.5)
filled_pauseslist[{time_s, time_e, words_idx}]one row per FP event
benchmark_{gender,speaker_id,power_status,age,orientation}_labelscalartask labels; nullable when the row isn't in that task
benchmark_{...}_splitstringtrain / dev / test per task

`alignments` — opt-in via load_dataset(repo, "alignments"). Join back on instance_id. Contains everything the viewer can't render cleanly.

ColumnTypeNotes
instance_idstringjoin key
wordslist[{time_s, time_e, char_s, char_e}]rough word-level char+time anchors
words_alignlist[{text, char_s, char_e, time_s, time_e, words_idx}]fine-grained forced-alignment
chars_alignlist[{text, time_s, time_e, char_s, char_e, words_align_idx}]flattened per-character alignment
primary_stresslist[{words_align_idx, stress, nuclei[], raw[]}]per-word stress annotation
linguistic_annotationlist[{words_idx, id, text, lemma, upos, xpos, feats, head, deprel, misc}]CoNLL-U style UD parses
align_textgrid, pause_textgrid, stress_textgridstring, nullablefull TextGrid file contents inlined

Quick example:

python
from datasets import load_dataset

ds = load_dataset("porupski/ParlaSpeech-HR-benchmark_v3", "default", split="train")
gender_train = ds.filter(lambda r: r["benchmark_gender_split"] == "train")
print(len(gender_train), gender_train[0]["speaker_gender"])

# Attach alignments when needed
align = load_dataset("porupski/ParlaSpeech-HR-benchmark_v3", "alignments", split="train")

Benchmark tasks

  • —Gender classification (M/F)
  • —Speaker identification (50-class)
  • —Power status classification (Coalition/Opposition)
  • —Age estimation (regression, years at recording)
  • —Political orientation (regression, −3 to +3 scale)

Use cases

  • —Speaker profiling and verification
  • —Filled pause detection and classification
  • —Primary stress frame prediction
  • —Fine-tuning Wav2Vec2 models for Slavic speech
  • —Benchmark evaluation for speech processing pipelines

References

  • —ParlaSpeech v3 paper: https://arxiv.org/abs/2511.01619v2
  • —ParlaSpeech landing page: https://clarinsi.github.io/parlaspeech/
  • —Full corpus: https://www.clarin.si/repository/xmlui/handle/11356/1833

License

CC BY-SA 4.0 (inherited from ParlaSpeech-HR).