PERIT-S2S · Public services

Benefits, permits and civic helplines, where callers read reference numbers aloud, badly.

Audio comes from municipal helplines, benefits intake, permit desks. Scored on task success, with a working public services operator as the control.

820
Clips in the target held-out set
63h
Audio, never published
9
Languages in the set
1
Working public services operator as control

Leaderboard

Illustrative figures. Public runs begin when the held-out sets close.
ModelTask success (%)Median turn latencyBarge-in recovery
Human control · sector agentHuman control
93.7 ± 2.2
505ms96.1%
Frontier realtime · lab BFrontier
60.2 ± 1.2
849ms65.7%
Frontier realtime · lab AFrontier
54.4 ± 0.8
676ms68.1%
Frontier multimodal · lab CFrontier
51.8 ± 2.3
1229ms59.6%
Cascaded stack (ASR → LLM → TTS)Specialist
46.6 ± 2.1
1310ms50.3%
Open weights · 8B speechOpen weights
27.2 ± 2.0
1258ms35%
Task families

What the set actually asks for.

  • Turn-taking & barge-in
  • Task completion
  • Instruction adherence
  • Disambiguation & repair
  • Escalation & refusal
  • Tone under pressure
Method & limits

How a run on this set is produced.

  1. 01Callers are scripted personas played by trained speakers, not synthesised — including the ones who talk over the agent.
  2. 02Every call is scored against the rubric by two reviewers; a third adjudicates splits.
  3. 038 calls per persona per model. Reported intervals are bootstrapped at 95%.
  4. 04Human control is a licensed agent from that sector handling the same personas blind.

Limits, stated up front: 9 languages only, single-channel audio, and no cross-sector transfer. The control is one operator per sector, so the human line carries its own error bar — it is a working standard, not a ceiling.