Does anyone have a tuning guide for the RTS/ASR in Breeze?

By tuning, I mean adjusting phrases, utterances, and confidence levels for a speech recognition snap-in.

The documentation is sparse ...

Hi, sorry for the very slow response here.

For RealTime Speech, there are APIs documented in the developer guide for setting confidence levels for various phrases. That is the only tuning available for that engine.

ASR functionality is provided by NICE or other MRCP speech engine. Any tuning is done directly on that speech engine rather than through Breeze or AAMS.

JoelEzell, thanks for the reply.

Can you point me to the Developer Guide that references confidence level tuning?

I found 'Avaya Engagement Designer Developer's Guide' v3.2 but it only lightly touches on confidence levels.

Sorry, this is the Realtime Speech SDK:
https://www.devconnectprogram.com/fileMedia/download/ccc2136f-b87f-466b-b1fc-a0d07b62d772

You could theoretically invoke these APIs from ED using the REST task, or you could use another application (even Postman) to invoke them.

Thank you.

Do you have any recommendations for tuning phrases? Something to start with?

When defining a search phrase with RTS, you define a confidence level. This confidence level determines the minimum confidence level the recognition engine must match on, before a successful speech search match can be returned. The confidence level range is between 1 and 100.

Setting the value too low can lead to false positive matches.

Setting the value too high can lead to missing real matches.

As per the programmers guide, the guidance would be to start at a relatively low level of confidence (e.g., 20), and update (via the REST APIs) based on testing. In many instances this will be best on manual test calls.

The programmers guide also provides details on how to write effective search phrases. RTS uses a phonetics based speech recognition engine, and is continuously monitoring an active conversation. Unlike a voice enabled IVR, which looks for a finite grammar within a finite period of time, the recognition engine is listening to the entire conversation. So searching for short phrase like "no", "yes", etc., will lead to a lot of false matches.

>> SeamusHayes - thank you.

The programmers guide you refer to, is this the same as developers guide JoelEzell cites?

Yes. In the documentation folder, there is a document called "Guidelines for Creating Effective Speech Search Queries"