Oasis Speech is the recognition service that turns a caller’s voice into text for every Dialogist agent. It runs on our servers, or on yours, and no caller audio goes to a third-party provider. Version 2 is live, and every agent now uses it.
Built for calls
Version 2 is an architectural change to the model, made with one goal: accuracy and quality of recognition on phone calls. We built it for narrowband audio, background noise, a caller who starts mid-word and, above all, short answers.
- Short replies. “Yes”, “Thursday”, a surname. The error rate on one-word replies fell by 49% (median across 48 languages), and every language improved.
- Silence stays empty. For a cough or a click on the line, Oasis Speech returns an empty result and does not invent a word.
- A confidence score that means something. The agent knows when it did not hear well, and asks again instead of guessing. On call-center audio the calibration error fell by 74%.
- Speed. Oasis Speech recognizes a short reply in about 150 ms.
48 languages
The agent now hears 48 languages. Each has a grade on the new Languages page, from A++ down, measured in phone quality. 37 languages are graded A++, and 46 of the 48 are A+ or better.
The agent also speaks all 48, so a caller can talk to it in any of them and hear the answer in the same language.
Eight languages up close
| Language | Error rate reduction |
|---|---|
| Serbian | 51% |
| Bosnian | 49% |
| Croatian | 44% |
| Italian | 40% |
| German | 36% |
| Spanish | 32% |
| Portuguese | 31% |
| French | 27% |
Read more
- Oasis Speech 2: recognition in 48 languages: the full per-language results and the method.
- The shortest answers are the hardest: why short replies are the hard case, and the path we built for them.