syvai/danish-diarization-bench

Danish Diarization Benchmark (Synthetic) — v2

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

What changed in v2 (2026-05-18)

• Per-segment text — each entry in segments now carries its text field directly. The redundant parallel texts column has been removed. Old consumers that joined segments[i] with texts[i] should switch to segments[i]["text"].
• Silent padding trimmed. Audio is now trimmed to [first_segment_start - 0.3 s, last_segment_end + 0.3 s] and segment timestamps are rewritten relative to the new start. The previous version intentionally added 1-5 s of leading/trailing silence to K=1 clips, but this caused autoregressive diarization models that emit speech every time they're asked to transcribe to hallucinate filler tokens (e.g. "Ja", "Mh") into the silent regions. Baselines below are re-measured on v2 audio.
• Schema: texts column removed, everything else unchanged. segments JSON is now [{start, end, speaker, text}, ...].

Baselines (DER, lower is better)

Evaluated on a 100-row sample with collar=0.0 s (no leniency), using pyannote.metrics.DiarizationErrorRate, audio capped at 30 s.

Model	Overall DER	K=1	K=2	K=3	K=4
syvai/hviske-v5.1-diarize-ts	15.19%	6.20%	11.44%	10.38%	25.99%

Reference-only model — hviske-v5.1-diarize-ts produces speaker tokens and 100 ms timestamps in a single auto-regressive decoder pass. The full prior pyannote / DiariZen baseline grid was measured on v1 audio; numbers will shift on v2 because most clips no longer have the 1-5 s silent padding.

Timestamp accuracy (matched segments, IoU ≥ 0.05):

• |Δstart| median 49 ms · p90 819 ms
• |Δend| median 71 ms · p90 1134 ms
• Segment IoU (matched) median 0.965

Composition

Rows	3996
Speakers per row	1 - 8 (~500 rows per K)
Audio	16 kHz mono WAV, trimmed to labeled region ±0.3 s
Total speech	~34 h
Source parquets	14 distinct from danish-asr-unified

Fields

Field	Type	Description
id	string	Sample identifier (sample_NNNNN)
audio	audio	16 kHz mono WAV bytes, trimmed
num_speakers	int32	Number of distinct speakers (1..8)
duration	float	Length in seconds (after trim)
segments	string (JSON)	List of {start, end, speaker, text}
sources	string (JSON)	Per-speaker source dataset (voxpopuli, nst_da, ...)

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"])
for s in segments:
    print(f"  [{s['start']:.2f} - {s['end']:.2f}]  {s['speaker']}  {s['text'][:80]!r}")

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%). Originally built 2026-05-15. v2 rebuilt 2026-05-18 with text inlined and silent padding trimmed.