syvai/danish-diarization-bench

Danish Diarization Benchmark (Synthetic)

A 3996-row synthetic speaker-diarization benchmark in Danish, built by mixing single-speaker utterances from syvai/danish-asr-unified into multi-speaker recordings.

Baselines (DER, lower is better)

Evaluated on a balanced 400-row subset (50 clips per speaker count K=1..8) with collar=0.0s (no leniency), using pyannote.metrics.DiarizationErrorRate. RTFx measured on a single NVIDIA RTX 3090. Each model uses its default pipeline configuration straight from the Hub - no Danish fine-tuning, no hyperparameter search.

Model	Overall DER	RTFx	K=1	K=2	K=3	K=4	K=5	K=6	K=7	K=8
pyannote/speaker-diarization-3.1	32.47%	15.7	17.5	26.2	27.7	26.9	31.5	33.8	35.5	38.7
pyannote/speaker-diarization-community-1	37.66%	16.4	17.4	31.2	39.1	35.6	42.8	38.9	39.6	39.2
BUT-FIT/diarizen-wavlm-large-s80-md-v2	41.00%	6.6	21.7	23.3	34.0	38.7	47.0	43.9	42.9	44.9
BUT-FIT/diarizen-wavlm-base-s80-md	41.47%	6.4	21.6	24.8	32.6	37.0	45.3	44.8	44.5	47.1
BUT-FIT/diarizen-wavlm-large-s80-mlc	61.57%	22.2	15.3	20.3	45.1	54.6	61.6	67.3	70.8	74.1

Notes

• DER is high across the board (vs the typical 12-20 % seen on AMI-SDM) because the source utterances are clean single-speaker ASR clips concatenated with abrupt cuts and synthetic overlap - very different from natural meeting acoustics. The benchmark is relative: it ranks models on Danish far-field-style audio, not an absolute estimate of meeting-room DER.
• DiariZen RTFx looks low because the short clips (mean 33.6 s) are dominated by per-clip pipeline overhead; on full meetings the same models hit RTFx 30-50.
• diarizen-wavlm-large-s80-mlc (multilingual) does worse than the English-focused v2 - the v2 checkpoint still produces better speaker representations even on unseen Danish.
• pyannote/speaker-diarization-3.1 wins by 5-9 pts over every DiariZen variant, consistent with pyannote's recent recipe being more robust to domain shift.

Composition

Rows	3996
Speakers per row	1 - 8 (~500 rows per K)
Overlap rate	7.9 % (AMI-like moderate)
Clip length	5.7 s - 76.6 s (mean 33.6 s)
Total speech	34 h
Source parquets	14 distinct from danish-asr-unified
Audio	16 kHz mono FLAC

Fields

Field	Type	Description
id	string	Sample identifier (sample_NNNNN)
audio	audio	16 kHz mono FLAC bytes
num_speakers	int32	Number of distinct speakers (1..8)
duration	float	Total length in seconds
segments	string (JSON)	Ground-truth: list of {start, end, speaker}
sources	string (JSON)	Per-speaker source dataset (voxpopuli, nst_da, ...)
texts	string (JSON)	Original transcripts per speaker

Usage

import json
from datasets import load_dataset

ds = load_dataset("syvai/danish-diarization-bench", split="test")
ex = ds[0]
segments = json.loads(ex["segments"])
print(ex["num_speakers"], ex["duration"], segments[:3])

Build provenance

For each synthetic row K speakers were sampled from K distinct parquet files in syvai/danish-asr-unified (when K <= pool size = 14). Each speaker contributes one 4-12 s random crop of its utterance, mixed in time with a 40 % overlap probability per transition (resulting overlap rate ~7.9 %). For K=1, the clip has uniformly random leading + trailing silence (1-5 s each).

Built 2026-05-15 from 14 randomly-sampled parquets of syvai/danish-asr-unified.