hours        2,000.24        mixtures   159,037
files          642 parquet   overlap    156.3 h (7.7% of audio)
speakers     4,660 derived identities
sessions     1-4 speakers, median 45 s
cameo        16.4% of mixtures contain a speaker who talks for 1-6 s in total

Each row carries audio (16 kHz PCM16), rttm, duration, num_speakers, turns, has_cameo, min_turns_per_speaker, overlap_sec. The RTTM is a construction, not an annotation — every turn was placed on the timeline by the generator, so it is correct by definition.

Why v2 exists — a duration leak in v1

In v1 each speaker received a fixed 4 turns, so session duration was a direct function of speaker count. Measured on the released v1 data:

1-spk  median 18.2 s   range 11.1-25.2
2-spk         30.4 s         23.0-36.6
3-spk         41.6 s         21.0-54.5
4-spk         54.6 s         38.0-73.0     <- 2-spk max 36.6 < 4-spk min 40.3

The 2-speaker and 4-speaker ranges do not overlap. A model can read the number of speakers off the length of the audio without listening to it — the same class of free shortcut as giving each speaker one contiguous block, which is why turns are interleaved in the first place. It would not transfer: a real consultation's length says nothing about how many people are in the room.

v2 budgets turns per session (12 ± 2) and splits them unevenly between speakers using random weights, so length no longer encodes speaker count:

1-spk  median 50.9 s   p5-p95 35.2-65.6
2-spk         46.5 s          35.1-57.6
3-spk         41.3 s          25.1-56.3
4-spk         41.2 s          28.4-55.2

Spread of medians fell from 36.4 s to 9.8 s and the ranges now overlap almost entirely. The residual 9.8 s is explained: cameo turns are 1–2 s and cameos concentrate in 3–4 speaker sessions, shortening them slightly.

The uneven split is deliberate beyond fixing the leak — a session might allocate turns [9,5] or [4,2,2,2], because no real consultation gives every participant equal airtime.

Cameo speakers — the measured failure this addresses

An earlier Sortformer fine-tune was scored against a held-out 4-speaker medical consultation and predicted 3 speakers, dropping one entirely. The dropped speaker talked for 5.20 s across 3 turns (2.00/1.84/1.36 s) — 9.9% of the DER denominator. Aggregate DER improved while this happened: only +3.8pp surfaced as missed speech, the rest being absorbed into the surviving speakers' slots, which corrupts their transcripts too.

The cause was in the data. The generator enforced ≥2 turns of ~4 s per speaker, so the least any training speaker ever spoke was ~8 s. A speaker who says one line had never been seen.

In v2, 16.4% of mixtures cast a cameo: 1–3 turns of 1–2 s each, excerpted at a random offset so cameos do not systematically begin at an utterance onset. Cameos are cast from speakers holding few clips — clustering produced 2,637 single-clip speakers, which looked like a defect and are in fact exactly right for a voice that must appear once and nowhere else.

The fix is short turns, not few turns. The majority of speakers still get ≥2 turns, which is what closes the positional shortcut.

Other design points, each from a measurement

Overlap is short. Median ~0.84 s, ~80% under 1.5 s. An earlier attempt at median 4.09 s was unintelligible to a native listener, and training an ASR on unrecoverable audio teaches hallucination.
Per-speaker level varies ±2.5 dB. Equal-RMS speakers make overlap maximally hard and unlike any real room.
Chunks are cut at silence. Boundaries are placed at low-energy points so a chunk does not begin mid-phoneme; a hard cut creates an artificial onset a diarizer can learn as a speaker-change cue that does not exist in real audio.
No speaker ever overlaps themselves — a hard clamp on each speaker's last end time, not merely turn ordering.
Single-speaker sessions are 29.5% of sessions. Without them a model learns that speech implies several speakers and over-segments a doctor's two-minute explanation into phantom speakers.

Replay factor

Drawn from 2,091.9 h of source at 0.96× — under one pass per epoch. The earlier fine-tune drew 400 h of mixtures from ≤159 h of source, a 2.7× replay, so its "3 epochs" were really ~8 passes over the same audio. Gains plateaued after epoch 1 with speaker confusion bit-identical between epochs 1 and 2.

Provenance and licence

Derived from DimitarV/eurospeech-bg-single-speaker, itself mined from EuroSpeech (Bulgarian National Assembly proceedings). Speaker identities are derived by clustering TitaNet-L embeddings — a label means "these clips share a voice", never a named person. See speaker_labels_v1.json in the source repo.

Source clips were verified single-speaker by Sortformer at mining time with a precision-favouring filter (98.1% keep rate over 497 shards).

See DATA.md for the full training recipe including the real-conversation rehearsal and probe sets.

Downloads last month: 176

Total file size:

183 GB