VoiceArena launches Diarization Bench: 12 systems, NVIDIA Nemotron 3 leads at 14.72% DER

rohanpaul_ai · x · 2026-09-24

VoiceArena launched Diarization Bench, scoring 12 speaker-diarization systems on 139 real-world conversations (rooms, calls, constant interruptions) with a single metric: diarization error rate. NVIDIA's Nemotron 3 leads at 14.72% DER — but with a 250 ms timing collar it scores 4.29%, revealing how sensitive results are to the tolerance window around speaker changes. Accuracy degrades sharply with overlapping speech and more speakers. It joins STT, TTS, and conversational-agent boards on VoiceArena.

Original post →

More from Research

Research channel →