united-nations/transcription-corpus

UN Transcription Corpus

Two splits of UN meeting audio paired with official verbatim records.

Splits

sessions — Whole meeting sessions (SC + GA plenary)

One row per meeting. Audio from UN Web TV, verbatim records from documents.un.org.

Column	Description
symbol	UN document symbol, e.g. S/PV.9826
webtv_url	URL on UN Web TV
duration_ms	Session duration in milliseconds
num_speakers	Number of speaker turns in the verbatim record
audio_floor	Floor audio (MP3)
audio_{lang}	Simultaneous interpretation per UN language (null if unavailable)
pv_{lang}	Official verbatim record text per language (null if unavailable)

Bodies covered: Security Council (S/PV.*), General Assembly plenary (A/{n}/PV.*)

gadebate — GA General Debate speeches

One row per country speech in the General Debate. Audio from UN Radio.

Column	Description
session	GA session number (70=2015 … 80=2025)
year	Calendar year
country_iso	ISO 3166-1 alpha-2 country code
country_name	Country name
original_lang	ISO 639-1 code of delivery language
speech_date	YYYY-MM-DD
audio_floor	Floor audio (null if original language is a UN language)
audio_{lang}	Simultaneous interpretation per UN language
orig_lang_text	As-delivered text from gadebate.un.org (null if unavailable or UN language)

Audio tracks

Each row has up to 7 audio tracks:

• audio_floor — floor/original-language audio
• audio_en/fr/es/ar/zh/ru — simultaneous interpretation into each UN official language

Audio files are stored separately and referenced by path in the Parquet files. To access audio directly:

https://huggingface.co/datasets/united-nations/transcription-corpus/resolve/main/data/sessions/audio/{symbol}_{lang}.mp3

Note on WER

Verbatim records are lightly edited. WER of 20-40% is expected even for high-quality ASR. For Chinese, use CER as the primary metric.

Attribution

Audio: United Nations. Reproduced under UN Terms of Use. Transcripts: United Nations Official Document System. Public domain.