Dataset Card for TRP-Bench-200 (update0909/proactive-turn-taking-bench-200)

1. Dataset Summary / 数据集简介

TRP-Bench-200 (Proactive Turn-Taking at Transition Relevance Places Benchmark) is a fine-grained millisecond-annotated evaluation benchmark designed for real-time full-duplex conversational voice agents. Curated by Yixin Zhang (张益新) at Yuansiwujie (源思无界), it evaluates whether a conversational agent can distinguish between mid-utterance thinking pauses (where speaking causes a false interruption) and Transition Relevance Places (TRPs) (where the speaker yields the floor or reaches a syntactic/prosodic boundary), while filtering out low-utility conversational filler via the Three-Question Validity Filter ($Q_1 \land Q_2 \land Q_3$).

• Total Multi-Turn Dialogues: 200 trajectories (diag_001 to diag_200)
• Intra-Turn Thinking Pauses: 480 annotated mid-clause pauses ($300\text{--}1800\text{ ms}$, is_trp = false)
• Ground-Truth TRPs: 600 annotated turn boundaries (is_trp = true), including 250 high-utility proactive speaking opportunities (PROACTIVE_SPEAK) and 350 deliberate silence moments (DELIBERATE_SILENCE)
• Domains Covered:
  1. technical_incident (Incident Response & SRE Failover — 50 dialogues)
  2. scheduling (Clinical Consultation & Multi-Constraint Scheduling — 50 dialogues)
  3. collaborative_debugging (Technical Pair Programming & Debugging — 50 dialogues)
  4. planning (Embodied Robot Co-Pilot & Mission Planning — 50 dialogues)

---

2. Dataset Schema / 数据结构说明

Each entry in benchmark_results_200.json (and data/dialogues_200.json) conforms to the JSON Schema in data/schema.json:

{
  "id": "diag_001",
  "trajectory_id": "diag_001",
  "domain": "technical_incident",
  "scenario_title": "Production Redis Cluster Failover & Architecture Review",
  "turns": [
    {
      "turn_id": 0,
      "speaker": "user",
      "text": "If the primary Redis node in us-east-2 crashes... we should failover immediately to the replica node in us-east-2.",
      "start_ms": 0,
      "end_ms": 3600,
      "is_trp": true,
      "acoustic_trp_prob": 0.94,
      "semantic_grounding_prob": 0.96,
      "is_thinking_pause": false,
      "pause_duration_ms": 1000,
      "events": [
        {
          "event_id": "diag_001_t0_e1",
          "timestamp_ms": 1800,
          "event_type": "thinking_pause",
          "text": "[650ms thinking pause: inspecting topology]",
          "duration_ms": 650,
          "is_trp": false,
          "acoustic_trp_prob": 0.08,
          "semantic_grounding_prob": 0.50,
          "candidate_response": {
            "text": "Yes, I am listening! What would you like to do next?",
            "sentence_count": 2,
            "category": "NOISE_FILLER",
            "q1_new_info": false,
            "q2_immediate_need": false,
            "q3_better_now": false,
            "net_utility_raw": 0.10,
            "optimal_action": "DELIBERATE_SILENCE"
          }
        },
        {
          "event_id": "diag_001_t0_e3",
          "timestamp_ms": 3600,
          "event_type": "trp",
          "text": "[Terminal boundary: terminal pitch fall -38Hz]",
          "duration_ms": 1000,
          "is_trp": true,
          "acoustic_trp_prob": 0.94,
          "semantic_grounding_prob": 0.96,
          "candidate_response": {
            "text": "Notice that the replica node is currently provisioned in us-west-1, not us-east-2. Failing over to it will introduce a 65 ms cross-region round-trip latency across all database writes.",
            "sentence_count": 2,
            "category": "PREMISE_CORRECTION",
            "q1_new_info": true,
            "q2_immediate_need": true,
            "q3_better_now": true,
            "net_utility_raw": 0.92,
            "optimal_action": "PROACTIVE_SPEAK"
          }
        }
      ]
    }
  ]
}

Field Definitions / 字段定义

Field	Type	Description
trajectory_id	string	Unique identifier (diag_001 .. diag_200)
domain	string	Task domain (technical_incident, scheduling, collaborative_debugging, planning)
events[].event_type	string	speech_segment, thinking_pause (mid-utterance pause), or trp (Transition Relevance Place)
events[].duration_ms	int	Duration of speech segment or silence window in milliseconds
events[].acoustic_trp_prob	float	$A(t) \in [0, 1]$: Acoustic-prosodic & syntactic completion probability
events[].semantic_grounding_prob	float	$S(t) \in [0, 1]$: Memory & speaker-grounded confidence
candidate_response.category	string	TASK_ADVANCEMENT, PREMISE_CORRECTION, BRIEF_CONFIRMATION, or NOISE_FILLER
candidate_response.q1_new_info	bool	$Q_1$: Provides novel information absent from shared working memory
candidate_response.q2_immediate_need	bool	$Q_2$: Immediately required to prevent an error or advance the current step
candidate_response.q3_better_now	bool	$Q_3$: Speaking now strictly dominates remaining silent
candidate_response.optimal_action	string	Ground-truth action (PROACTIVE_SPEAK or DELIBERATE_SILENCE)

---

3. Baseline Benchmark Results / 基准评测结果

Model / System	TTL (ms) $\downarrow$	FIR (%) $\downarrow$	NFR (%) $\downarrow$	Annoyance (1–7) $\downarrow$	Helpfulness (1–7) $\uparrow$
Method A ($\tau = 300\text{ ms}$ Fixed VAD)	450.0	79.38%	74.52%	6.97	1.20
Method A ($\tau = 600\text{ ms}$ Fixed VAD)	750.0	45.42%	69.44%	6.39	1.20
Method A ($\tau = 800\text{ ms}$ Fixed VAD)	950.0	15.00%	64.73%	5.23	1.20
Method A ($\tau = 1200\text{ ms}$ Fixed VAD)	1350.0	4.38%	95.83%	6.63	1.20
Method B (Semantic VAD, $A \times S$)	200.0	0.00%	55.20%	2.82	2.49
Method C (Ours: $U = A \times S \times N, \theta=0.50$)	200.0	0.00%	0.00%	1.19	6.35

---

4. Citation & Author Metadata / 引用与作者信息

• Author: Yixin Zhang (张益新)
• Affiliation: Yuansiwujie (源思无界)
• ORCID: https://orcid.org/0009-0004-7641-4450
• Homepage: https://yixinzhangagent.github.io
• GitHub Repository: https://github.com/yixinzhangagent/proactive-turn-taking-benchmark
• GDE Profile: https://g.dev/updatezhang

@article{zhang2026proactive,
  title     = {Proactive Turn-Taking at Transition Relevance Places for Real-Time Conversational Agents},
  author    = {Zhang, Yixin},
  year      = {2026},
  journal   = {arXiv preprint},
  note      = {Affiliation: Yuansiwujie. ORCID: 0009-0004-7641-4450},
  url       = {https://github.com/yixinzhangagent/proactive-turn-taking-benchmark}
}