Every voice AI demo sounds brilliant because the demonstrator speaks slowly, clearly and in a quiet room. Real callers do none of those things. They ring from a car, they talk fast, they have accents the training data under-represents, and they say place names that trip up every speech model ever built.
The demo voice versus the real caller
Recognition accuracy measured on clean audio tells you very little. What matters is accuracy on your calls: your customers, your destinations, your reference formats. We benchmark every release against a corpus of real, consented calls that is deliberately heavy on strong regional accents and poor line quality.
Where recognition breaks
- Place names and resort names, especially non-English ones said with a regional accent
- Postcodes and booking references — short strings with no linguistic context
- Numbers: 'fifteen' and 'fifty' remain the classic failure
- Dialect words and local phrasing for dates and times
How we test
Beyond the accent corpus we simulate bad lines: compression artefacts, dropouts and background noise layered over clean recordings. Each scenario has a target for recognition and, more importantly, for task completion — because a system that mishears a word but confirms it back and recovers has still done its job.
Design for recovery, not perfection
No recogniser is perfect, so the conversation design has to assume mistakes. Confirm back anything that matters, offer to spell or use the keypad for references, and never let a single misheard word send the call off a cliff. Recovery is a feature; it should be designed as carefully as the happy path.
A voice agent that only understands the demo voice has not met your customers yet.
