Multi-speaker YouTube conversations from YODAS2, cut into chunks of up to 10 minutes and transcribed speaker by speaker with DiCoW v3.3 conditioned on an automatic diarization. These are the statistics of the 4.6k hours that Pandemonium was trained on.
Because of possible licensing issues with the source recordings, neither the audio nor the labels are shared. The corpus is released as the scripts that mine it, plus a list of 100 sample recordings, so that it can be rebuilt. The public recipe uses Nemotron-3-Diarization and DiCoW v3.3 and gives a comparable corpus, not an identical one.
Recordings were drawn at random, stratified by language, from those with at least 2 speakers, 5 s of overlapped speech and 15 s from the quietest speaker. Chunks with more than 10 speakers were dropped; the labels identify at most 8. The last step keeps the languages Parakeet-v3 supports, the only filter applied to Pandemonium's training data.
Hours of audio by the number of speakers the diarization found in a chunk.
Number of chunks by length in minutes. Recordings are cut at silences near every 10 minutes, so most chunks are full length and the rest are recording tails.
The paper's DiCoW study (Table 2) also tested three label filters. They are in the recipe as optional steps; Pandemonium was trained without them. Run on this corpus, the two rule-based ones mark:
A flagged chunk keeps its unflagged windows of at least 2 minutes; an unsalvageable one has none. The third filter, an LLM language check, needs an API and is not counted here.