README.md

metadata

license: cc-by-nc-4.0
language:
  - zh
language_bcp47:
  - zh-TW
task_categories:
  - automatic-speech-recognition
  - text-to-speech
  - audio-classification
pretty_name: Parliament
tags:
  - Taiwan
  - Traditional Chinese
  - parliament
  - speech
  - ASR
  - TTS
  - Legislative Yuan
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet

Parliament

Parliament is an OpenFormosa Traditional Chinese speech dataset derived from the zh_tw split of disco-eth/WorldSpeech. It contains audio clips and human transcripts from Taiwan Legislative Yuan IVOD parliamentary proceedings.

This release keeps only rows that passed the Taiwan-OmniData / FineWeb2-style text filtering pipeline. Audio is preserved from the upstream dataset and cast as a Hugging Face Audio(sampling_rate=24000) feature.

Dataset Summary

• Dataset repo: OpenFormosa/parliament
• Dataset name: parliament
• Credit: OpenFormosa
• Source dataset: disco-eth/WorldSpeech
• Source config: zh_tw
• Source split: train
• Upstream source: Taiwan Legislative Yuan IVOD parliamentary proceedings
• Audio: included, Opus/Ogg bytes, exposed as datasets.Audio
• Text: human transcript
• Rows after filtering: 308,521
• Rows rejected by filtering: 24,762
• Estimated text tokens after filtering: 11,464,559
• License: CC BY-NC 4.0, following the upstream WorldSpeech dataset card

Fields

• audio: audio clip, Hugging Face Audio(sampling_rate=24000)
• text: filtered human transcript
• human_transcript: original human transcript from WorldSpeech
• asr_transcript: upstream ASR transcript
• cer: upstream character error rate signal
• snr: upstream signal-to-noise estimate
• dnsmos_sig, dnsmos_bak, dnsmos_ovr, dnsmos_p808: upstream DNSMOS quality signals
• duration: clip duration in seconds
• source: upstream source label
• source_url: original IVOD media URL
• source_start_s, source_end_s: segment offsets in the original media
• session_date: parliamentary session date if available
• segment_id: upstream segment id
• doc_id: Taiwan-OmniData document id
• row_index: row index in disco-eth/WorldSpeech, config zh_tw, split train
• estimated_tokens: heuristic token estimate for the transcript
• quality_score, taiwan_relevance_score, zh_tw_naturalness_score, safety_score, pii_risk_score: filtering/scoring signals
• filter_metadata: compact provenance metadata for the filtering pass

Filtering

The source rows were processed through the Taiwan-OmniData filtering pipeline:

• text normalization and metadata preservation
• Taiwan/Traditional Chinese locale tagging
• FineWeb2-style repetition and boilerplate filters
• safety and PII-risk scoring
• train/eval separation metadata

Rejected rows are not included in this release. Top rejection reasons from the filtering run:

• fineweb2_top_2gram_fraction: 8,336
• fineweb2_top_4gram_fraction: 5,966
• fineweb2_dup_5gram_fraction: 5,467
• fineweb2_top_3gram_fraction: 4,895
• fineweb2_dup_6gram_fraction: 95
• fineweb2_max_avg_word_length: 2
• fineweb2_dup_7gram_fraction: 1

Usage

from datasets import load_dataset, Audio

ds = load_dataset("OpenFormosa/parliament", split="train")
ds = ds.cast_column("audio", Audio(sampling_rate=24000))
sample = ds[0]
print(sample["text"])
print(sample["audio"]["sampling_rate"])

For metadata-only inspection without decoding audio:

from datasets import load_dataset, Audio

ds = load_dataset("OpenFormosa/parliament", split="train")
ds = ds.cast_column("audio", Audio(sampling_rate=24000, decode=False))
print(ds[0]["audio"].keys())

License And Attribution

This dataset follows the upstream WorldSpeech license declaration, CC BY-NC 4.0. Please also review the upstream WorldSpeech dataset card and the applicable terms for Taiwan Legislative Yuan IVOD content before commercial use.

If you use this dataset, please credit OpenFormosa and cite or link both the upstream WorldSpeech dataset and this derived dataset:

• OpenFormosa Parliament: OpenFormosa/parliament
• Upstream source dataset: disco-eth/WorldSpeech

Provenance

Generated by Taiwan-OmniData-R1 on 2026-06-06T18:31:39.990842+00:00. Audio was not re-encoded; the original WorldSpeech audio bytes were preserved and cast to the Hugging Face Audio feature.