This eight-hour pilot follows the Hugging Face structure used by ParlaSpeech-style speech corpora, with a compact schema tailored to Lok Sabha data. The default configuration contains one accepted aligned audio segment per row, with embedded 16 kHz audio, verbatim ASR, an explicitly separate edited UCR passage, word timings, speaker metadata, and source-order fields.

Configurations

default: 1,195 accepted audio segments (8.013 hours).
speakers: metadata for 111 MPs.
speech_sections: 193 longer ordered speech sections.
chunk_inventory: all 1508 evaluated chunks, including rejected gaps.

text contains the selected verbatim ASR transcription. text_normalised is its whitespace-normalised counterpart for ParlaSpeech-style tooling. text_ucr_edited contains the matched passage from the official Uncorrected Debate record; it is edited parliamentary text and must not be treated as a verbatim transcription. The nested words records follow ParlaSpeech's char_s, char_e, time_s, and time_e convention and add section-level and source-recording times.

Rows remain in parliamentary source order. speech_section_id and chunk_sequence group consecutive pieces of a longer speech. Rejected material is absent from the default audio configuration but remains explicit in the chunk_inventory configuration; continuity flags prevent silent concatenation across those gaps.

When generated by src/annotate_lssec.py, separate auxiliary configurations provide a deletion-only readable text view, timed disfluency events, lemmas and POS/dependency analyses linked to the aligned ASR words, experimental binary sentiment scores, and transparent lexical CAP topic candidates. These are machine annotations, not corrections to the ASR or human gold labels.

License and source terms

This test release is licensed under CC BY-NC 4.0. The underlying recordings and official records come from the Parliament of India (sansad.in and videolibrary.sansad.in). The original parliamentary source terms continue to apply; where they are more restrictive, those source terms govern.

Important limitations

Acceptance is based on conservative automatic UCR–ASR lexical alignment and has not yet been replaced by complete human transcription validation.
The pilot retains fixed 25-second chunk boundaries, which can split words.
Party, state, and constituency values are source snapshots and may not be historically correct for every sitting date.
Ninety accepted rows use prompted Whisper large-v3 rescue; the remaining 1,105 use the fine-tuned verbatim Qwen3-ASR model. asr_model identifies them.
This repository structure is compatible with ParlaSpeech-style use on the Hugging Face Hub; it does not claim formal ParlaMint XML conformance.

Loading

from datasets import load_dataset

speech = load_dataset("stcoats/Lok_Sabha_test", split="train")
speakers = load_dataset("stcoats/Lok_Sabha_test", "speakers", split="train")

Downloads last month: 1

Total file size:

551 MB