Your voice agent sounds right. Did it place the right order?

Description and use

IndicOrderBench detected a mismatch between the spoken and committed order in a voice agent, preventing customer impact.

Problem

Voice ordering agents may misinterpret or miscommit orders, especially with code‑mixed speech, leading to incorrect backend state.

Outcome

IndicOrderBench detected the mismatch between committed and expected order, causing CI failure before customer impact.

Workflow

• Customer speaks request in Hinglish
• Agent parses and confirms order
• Agent commits order to backend
• Backend state is compared to expected state

Evidence

The agent read back one wrap without onion. This is what it committed:

Source

A transcript-based test, or an LLM judge reading the conversation, passes this call. The customer gets two wraps with onion.

Source

IndicOrderBench is an open-source benchmark and CI gate for voice ordering agents.

Source