MESH - Egocentric Dynamic Transcription: leaderboard

MESH tests spatial hearing across synchronized receivers and moving speakers. Direction, distance, position and transcription tasks remain separate. Word error rate can exceed 100%. Overall gives equal weight to localization and per-question transcription quality; it is not an average of the printed task metrics. This board reports corpus WER (%). Lower is better.

Metric: Corpus WER (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2027. 10 models tracked. Updated 14 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Qwen 3.5 Omni Plus (Harness: MESH FOA-sep)73.1
2Qwen3.8 Omni Flash73.32#115
3Qwen 3.5 Omni Plus (Harness: MESH FOA-W)78
4MESH-agentEvaluation run: V2; Qwen3-14B planner and source specialist tools78.98
5MiDasheng-Spatial86.56
6BAT99.82
7JAEGER Classical100.75
8JAEGER Neural101.2
9Qwen3 Omni 30B A3B Instruct (Harness: MESH FOA-sep)123.7
10Qwen3 Omni 30B A3B Instruct (Harness: MESH FOA-W)128

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=mesh-egocentric-dynamic-transcription · How It Works · Data refreshed daily, snapshot 2026-11-16.