SmeanAI/diarization-dataset

Smean diarization training data

The training data of smean-diarizer-1 and smean-diarizer-large: about 1,190 h of 16 kHz mono audio with speaker-diarization labels (RTTM). Half is real meetings and broadcasts, half is simulated conversation, mostly in Khmer. Each training batch drew from the sets with the weights below.

Set	Train	Dev	Batch weight	What it is
real	1,155 recordings, 351.6 h	144, 35.1 h	0.50	AMI, AISHELL-4, AliMeeting, NOTSOFAR-1, VoxConverse (single channel)
sim_khmer_meetings	12,000, 417.3 h	150, 5.2 h	0.20	simulated Khmer meetings
sim_khmer_broadcast	6,000, 209.7 h	150, 5.3 h	0.15	simulated Khmer broadcast-style conversations
sim_multi_meetings	6,000, 206.7 h	150, 5.2 h	0.15	simulated multilingual meetings (English, Mandarin)

The real training recordings are AMI (134), AISHELL-4 (173), AliMeeting (209), NOTSOFAR-1 (465) and VoxConverse (174).

Layout

<set>/{train,dev}/wav.scp        recording id -> audio path (relative to the repo root)
<set>/{train,dev}/rttm           labels to use (see below)
<set>/{train,dev}/rttm_original  labels as first built
<set>/{train,dev}/all.uem        scored region of each recording (the whole file)
<set>/audio/...                  16 kHz mono WAV

Labels

Two label versions are included for every set:

• rttm — the current labels.
  • real: the corpora's reference labels with silence removed. These labels marked many pauses as speech (about 31% of AMI and 25% of AISHELL-4 label time was silence). A frame is kept only if Silero VAD hears speech or its energy is at least 6 dB above the recording's noise floor; gaps under 0.1 s are bridged. Overlap and speaker identities are unchanged.
  • sim_*: rebuilt from each source utterance with the same rule, so more pauses inside a turn are marked as silence: short pauses (< 0.5 s) per speech-minute go from about 1 to 3.6 (Khmer) and 5.8 (multilingual). The source speech is read sentences, which pause rarely, so the change is modest. The energy test also widens segment edges slightly: about 2.5% of the rebuilt label time lies outside rttm_original and 0.3% of the original label time is removed. Neither version has been compared in training yet.
• rttm_original — the labels smean-diarizer-1 and smean-diarizer-large were trained on: the corpora's reference labels for real; FastMSS output for sim_*, which merges same-speaker gaps up to 0.2–0.3 s.

How the simulated sets were made

Built with FastMSS from single-speaker speech:

• Khmer: Khmer ASR Cultural Dataset and Khmer ASR dataset by Digital Divide Data, and OpenSLR 42. The train sets use 27 speakers. The dev sets use 4 different speakers, held out of all training simulation.
• Multilingual: LibriSpeech and AISHELL-1, 1,336 speakers. Dev speakers are held out.
• Turn-taking: an HMM fitted on real meetings (meetings sets) and on VoxConverse broadcasts (broadcast set).
• Augmentation: room acoustics from 500 simulated rooms (pyroomacoustics) and MUSAN noise and music.

Augmentation at training time (extra noise, reverberation, telephone band, gain) is not baked into the audio.

Not included

The Khmer evaluation set (khmer-diarization-eval) was never used for training and is not part of this dataset.

Licences

This dataset combines material under these licences; use must respect each of them:

• AMI: CC BY 4.0
• AISHELL-4: CC BY-SA 4.0
• AliMeeting: CC BY-SA 4.0
• NOTSOFAR-1: as published by its authors
• VoxConverse: CC BY 4.0
• Khmer ASR Cultural Dataset and Khmer ASR dataset by Digital Divide Data: CC BY-SA 4.0
• OpenSLR 42: CC BY-SA 4.0
• LibriSpeech: CC BY 4.0
• AISHELL-1: Apache 2.0
• MUSAN: CC BY 4.0 and public-domain parts