Development of AI Live Customer Sentiment Analysis System
Live Sentiment Analysis continuously evaluates the emotional state of the customer during the call and displays an indicator on the operator's or supervisor's screen. Allows responding to sentiment changes before the customer hangs up.
Approaches to Real-Time Sentiment Analysis
Text analysis of transcript (delay 300–500 ms):
- Use transformer models like RuBERT for Russian
- Classify sentiment as POSITIVE | NEGATIVE | NEUTRAL
- Output confidence scores
Acoustic voice analysis (no transcription, delay <100 ms):
- Extract prosodic features: F0 (pitch), tempo, energy, jitter, shimmer
- Classify emotion without text: angry, sad, neutral, happy
Fusion: text + acoustics
- High energy + negative text = anger
- Low energy + negative text = distress
- Combine confidence scores
Timeline: text sentiment in real-time — 2–3 weeks. Acoustic + fusion — 4–6 weeks.