PERIT-S2S · Banking

Card disputes, KYC checks and payment failures, where a misheard digit is the whole error.

Audio comes from branch calls, fraud desks, collections. Scored on task success, with a working banking operator as the control.

1,180
Clips in the target held-out set
88h
Audio, never published
6
Languages in the set
1
Working banking 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
90.0 ± 1.0
565ms97.9%
Frontier realtime · lab AFrontier
66.9 ± 1.7
897ms70.1%
Frontier realtime · lab BFrontier
57.3 ± 2.3
811ms55.4%
Frontier multimodal · lab CFrontier
54.2 ± 1.8
1138ms58.1%
Cascaded stack (ASR → LLM → TTS)Specialist
50.0 ± 1.1
1499ms40%
Open weights · 8B speechOpen weights
29.4 ± 1.6
1348ms32.4%
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: 6 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.