we are currently testing the Genesys Cloud Agent Assist feature in our production environment. We are seeing very poor transcription accuracy for our agents who are working in noisy environments or using low-quality headsets. The AI suggestions are often irrelevant because the transcription engine is misinterpreting several key technical terms. Is there a way to ‘train’ the Agent Assist transcription engine with a custom dictionary or a language model for our specific industry?
Greetings Fio31. I am a compliance analyst and I have seen this accuracy problem impact our recording requirements. You cannot directly ‘train’ the base transcription model, but you can use ‘Speech and Text Analytics’ (STA) to improve the results. You can define a ‘Dictionary’ of terms in the Admin menu. This tells the engine to prioritize your specific industry terms when it hears something phonetically similar. It is not perfect, but it definitely helps the Agent Assist logic provide better suggestions.
I have helped several clients with this. To follow up on Eli93, you should also ensure that your agents are using ‘High Definition’ (G.711) audio for their interaction. If your SIP trunks are using a compressed codec like G.729, the transcription engine will struggle regardless of your custom dictionary. Also, the Agent Assist suggestions rely heavily on the ‘Knowledge Base’ articles. If your articles are not written in a way that matches the customer utterances, the AI will always provide poor suggestions.