{
 "body": "## What it measures\n\nM3-DuplexBench evaluates full-duplex spoken dialogue systems that can listen while speaking. It targets smooth turn-taking, backchannel handling, and user barge-in handling, capabilities that ordinary turn-based speech tests do not capture.\n\nThe benchmark supports English and Japanese, casual conversation, and multi-turn question answering. It also varies dialogue context so researchers can separate language, domain, and history effects.\n\n## How it is scored\n\nThe paper compares recent full-duplex systems under single-turn, user-only, and teacher-forced full-context conditions. It reports model-specific turn-taking characteristics and performance gaps across languages and domains. The abstract does not specify a single metric name, maximum, or baseline, so those values remain unknown here. Context mode must accompany every score.\n\n## Dataset and licence\n\nThe paper establishes the supported languages, dialogue types, and context conditions but does not state a total item count or licence in the abstract. Public test visibility is therefore unknown. Reproduction should use the authors\u2019 release and record audio provenance, transcript policy, and context condition.\n\n## Who publishes it\n\nRyo Fukuda, Atsushi Ando, Hiroki Kanagawa, Takatomo Kano, Marc Delcroix, Naohiro Tawara, and Yuya Chiba introduced the benchmark in a July 2026 arXiv submission to SLT 2026. No maintained public leaderboard is identified.\n\n## Lineage\n\nM3-DuplexBench is a standalone full-duplex spoken-dialogue benchmark. It addresses limited multilingual and multidomain coverage in earlier dialogue tests, but the paper does not name a specific predecessor or successor.\n\n## Saturation and contamination\n\nThe paper reports clear gaps across languages and domains and mixed effects of dialogue context. This indicates an open evaluation space. The source does not establish training-data contamination, so risk is unknown.\n\n## How to run it\n\nUse the same audio interface and context condition described by the paper. Report language, dialogue domain, whether the user can interrupt, latency or turn-taking settings, transcript handling, and model sampling. Do not combine context conditions into one score without showing the breakdown.\n\n## Reading the numbers\n\nA strong score indicates better performance in the tested dialogue setup. It does not prove natural turn-taking in unseen languages or noisy real conversations. Compare language and domain slices, and inspect interruptions and backchannels. Context mode is part of the result, not an incidental implementation detail.\n",
 "build": {
  "built_at": "2026-09-09T16:56:50+00:00",
  "commit": "0a599558854c0e238c03a0f0d725239cb28f9d11",
  "eligibility_as_of": "2026-09-09"
 },
 "disposition": {
  "canonical_id": "m3_duplexbench",
  "reasons": [],
  "status": "unassessed",
  "verified_results": []
 },
 "models_covered": [],
 "page": {
  "category": "generation",
  "contamination": {
   "note": "Training exposure is not established by the paper abstract.",
   "risk": "unknown"
  },
  "dataset": {
   "languages": [
    "English",
    "Japanese"
   ],
   "modalities": [
    "audio"
   ],
   "public_test_set": null
  },
  "freshness": {
   "researched": "2026-09-08",
   "researched_by": "GPT-5.6 Luna, luna-stream-a-004 (Codex coordinated)",
   "reviewed": "",
   "reviewed_by": ""
  },
  "harness": {
   "other": "The evaluation protocol described in the M3-DuplexBench paper."
  },
  "id": "m3_duplexbench",
  "measures": "M3-DuplexBench tests spoken dialogue systems that listen while speaking. It covers English and Japanese, casual conversation and multi-turn question answering, and multiple context conditions for studying dialogue history.",
  "metric": {
   "direction": "higher_is_better",
   "name": "dialogue performance",
   "unit": "score"
  },
  "name": "M3-DuplexBench",
  "page_kind": "benchmark",
  "paper": {
   "arxiv": "2607.29125",
   "title": "M3-DuplexBench: A Multi-Turn, Multilingual, Multidomain Benchmark for Full-Duplex Spoken Dialogue Models",
   "url": "https://arxiv.org/abs/2607.29125",
   "year": 2026
  },
  "publisher": {
   "authors": [
    "Ryo Fukuda",
    "Atsushi Ando",
    "Hiroki Kanagawa",
    "Takatomo Kano",
    "Marc Delcroix",
    "Naohiro Tawara",
    "Yuya Chiba"
   ],
   "org": "M3-DuplexBench authors",
   "url": "https://arxiv.org/abs/2607.29125"
  },
  "released": "2026-07",
  "saturation": {
   "note": "The paper reports language-, domain-, and model-specific gaps.",
   "status": "open"
  },
  "sources": [
   {
    "accessed": "2026-09-08",
    "title": "M3-DuplexBench paper",
    "url": "https://arxiv.org/abs/2607.29125"
   }
  ],
  "subcategory": "spoken dialogue",
  "summary": "M3-DuplexBench evaluates multilingual full-duplex spoken dialogue with turn-taking, backchannels, and user interruptions.",
  "tags": [
   "speech",
   "duplex",
   "multilingual",
   "turn-taking"
  ],
  "task_format": "Multi-turn audio dialogue with full-duplex interaction and context variants."
 }
}