Blog
Product updates, launches and use cases — written by the people who build and sell it.

A customer picks up, says labdien, and the model listening decides it just heard English — on 60% of real Latvian phone speech in our internal tests, from OpenAI's Whisper, the most widely deployed open speech model, against ground truth labelled by native speakers. The standard synthetic evaluation didn't just inflate the numbers; it picked the worst model on our board as the best one.

A bank announced a cashback card. I built the outbound campaign for it in Convershake in under half an hour: the flow, the questions it may answer, the outcomes it records. Nothing was dialled. Here is what I built, and what people said when I showed it.

Today we're introducing test calls from the script editor: type your phone number, press one button, and your AI voice agent calls you, running the exact script you have open on screen. You talk to it the way a borrower would, on a real phone line, and the conversation streams back to your screen while you're still on the call.

Here is what we shipped in Convershake in July. Three improvements your team will feel first — AI voice agents that handle borrowers answering in another language, campaign reruns with same-day redial control, and a safe way to retire outdated call scripts (the conversations your agents follow) — plus everything else, including call analysis in 14 more languages.

Here is what we shipped in Convershake in June. Three improvements your team will feel first — collection scripts that carry the fundamentals, a finish estimate before you launch, and safe answers to off-script questions — plus everything else, listed at the end.

Here is what we shipped in May. Four changes your collections team will feel first — plus everything else, listed at the end.

Here is what we shipped in April. Four big changes your team will feel first — more natural conversations, your own telephony, live call monitoring and two-channel recordings — plus everything else, listed at the end.