VoiceArena Diarization Bench - cpWER: leaderboard

A transcription system writes a speaker-attributed transcript of each of 139 recorded English conversations (2 to 8 speakers, 115 in person and 24 online, 272,690 reference words), and the number is concatenated minimum-permutation word error rate in percent. Unlike plain word error rate boards such as Open ASR, a word attributed to the wrong speaker counts as an error.

Metric: cpWER on 139 conversations (%). Source: voicearena.com. Saturation forecast: Rough model projection: around 2026. 12 models tracked. Updated 16 Nov 2026.

Top models

#ModelScoreOverall rank
1AirCaps A5S Async v123.89
2Nemotron 3.5 Transcribe32.43
3MOSS-Transcribe-Diarize34.73
4Scribe v235.73
5Muse Voice Transcribe36.53
6AssemblyAI Universal-3.6 Pro47.81
7Soniox v5 Async48.28
8Smallest AI Pulse50.85
9Voxtral Small Transcribe 359.65
10Voxtral Mini Transcribe 365.04
11Speechmatics Enhanced66.07
12Deepgram Nova 371.89

Interactive version: theaggregate.ai/benchmark?slug=voicearena-diarization-bench-cpwer · How It Works · Data refreshed daily, snapshot 2026-11-16.