CoolFace
Datasetpublic

Trelis/tiron-eval-meetings

Tiron evaluation meetings The 17 held-out whole meetings used for the benchmarks on the Trelis/tiron model card — far-field single-channel audio (16 kHz mono WAV) with reference speaker-attributed transcripts, packaged so results can be reproduced with the Tiron harness. split meetings source ami ES2004a, IS1009a, TS3003a, EN2002a AMI Meeting Corpus, single distant microphone (Array1-01) icsi Bmr013, Bmr018, Bro021 ICSI Meeting Corpus, mean of 4 distant PZM room… See the full description on the dataset page: https://huggingface.co/datasets/Trelis/tiron-eval-meetings.

sourceHugging Facecc-by-4.0updated 2mo agoView on Hugging Face
0likes90downloads
Dataset Card

Tiron evaluation meetings

The 17 held-out whole meetings used for the benchmarks on the Trelis/tiron model card — far-field single-channel audio (16 kHz mono WAV) with reference speaker-attributed transcripts, packaged so results can be reproduced with the Tiron harness.

splitmeetingssource
amiES2004a, IS1009a, TS3003a, EN2002aAMI Meeting Corpus, single distant microphone (Array1-01)
icsiBmr013, Bmr018, Bro021ICSI Meeting Corpus, mean of 4 distant PZM room microphones
notsofar10 MTG_* sessionsNOTSOFAR-1 eval set, one single-channel distant device per meeting

Each row: corpus, meeting_id, audio (whole-meeting 16 kHz mono WAV), utterances_json (list of {speaker_id, begin_time, end_time, text} reference utterances on the same timeline), unknown_spans_json (NOTSOFAR only: [start, end] spans of annotator-<UNKNOWN/> stretches, used by the masking rule), duration_s.

Licenses and attribution (all CC BY 4.0)

  • AMI Meeting Corpus — © University of Edinburgh and the AMI consortium, CC BY 4.0. Carletta et al., The AMI meeting corpus: a pre-announcement (MLMI 2005). Audio unmodified (Array1-01 distant channel); transcripts converted from the NXT annotations to a flat utterance list.
  • ICSI Meeting Corpus — © International Computer Science Institute, CC BY 4.0. Janin et al., The ICSI meeting corpus (ICASSP 2003). Change note: audio is a mean-downmix of four distant PZM room-microphone channels (chanE/chanF/chan6/chan7); transcripts converted from MRT.
  • NOTSOFAR-1 — © Microsoft, CC BY 4.0. Vinnikov et al., NOTSOFAR-1 Challenge (2024). Audio unmodified (one single-channel distant device per session, chosen deterministically); references from gt_transcription.json.

Scoring convention used on the model card (cpWER, pooled per corpus, and the NOTSOFAR <UNKNOWN/> masking rule) is implemented in the harness repo's eval scripts.