Call-Centre Speech Analytics

A call-analysis system that turns a call recording into who spoke when, what was said, how the customer felt and what the call was about. Speakers are separated with voice activity detection, ECAPA-TDNN speaker embeddings and clustering, speech is transcribed with Whisper, every customer turn gets one of six emotions, and BERTopic names the topic of the call. A quality analyst sees an escalation risk, the customer's emotion trend, talk share and silence. A 28-second call is analysed in about 4 seconds on a laptop CPU.

Category: Speech Processing / Speech Emotion Recognition / Speaker Diarization / Topic Modelling / Natural Language Processing

CSE Project 32 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.