Malayalam, Hindi and Tamil among others, for both speech recognition and speech synthesis, on the public telephone network — not an English model with an accent setting, and not text-only. This is the difference that decides most deals we win: a platform built for English-speaking markets can be excellent and still be unusable for a call centre in Kerala.

Parloa is an enterprise programme. Mirakash is configured and on a live number the same day, which changes what you can find out before committing — you evaluate a real call rather than a reference architecture.

Speech-to-text, the language model and text-to-speech are separate stages behind one interface, so any of the three can be swapped — for cost, for quality, or for a language a particular vendor handles badly. Most platforms in this category are a single opinionated stack. That is simpler, and it is also why a customer who needs a different speech vendor for one language has no move to make.

Calls bill on connected seconds rather than rounded-up minutes, and an answering machine is detected so a voicemail is neither billed nor talked to as though it were a person. On an outbound campaign where a third of dials reach voicemail, the difference between those two billing models is most of the bill.