Nvidia reports a 14.72 per cent error rate on a public leaderboard against 19.3 for the next system. The figures are the company's own.
Benchmark
VoiceArena Diarization-Bench
How accurately a system separates a recorded conversation into who spoke when, across a fixed set of multi-speaker English recordings.
- Maintainer
- VoiceArena
- Score
- % DER
- Last checked
- September 27, 2026
Leaderboard
No verified results recorded yet.
Methodology
Systems are scored by diarization error rate, the share of audio time assigned to the wrong speaker, missed entirely or wrongly marked as speech. The leaderboard's English set covers 139 conversations totalling roughly 22 hours. A lower figure is better.
Important limitations
The English set is one language and one collection of recordings, and a leaderboard position says nothing about performance on other languages, on noisier audio, or on conversations with more simultaneous speakers than the set contains. Submitted results are reported by the systems' own developers.