Krisp VIVA 2.0

### TL;DR

Krisp has launched VIVA 2.0, a server-side voice infrastructure SDK that transforms voice AI agents from reactive to predictive. It sits in the audio pipeline before speech-to-text, cleaning audio with Voice Isolation v3 and adding Turn Prediction v3, Interruption Prediction v1, and Signal Detectors. Already running inside Daily, Vapi, LiveKit, Vodex, and Ultravox, VIVA processes over 10 billion voice AI minutes per year.

Key Insights & Metrics

Key Features

• Voice Isolation v3: ground-up rebuild that isolates the primary speaker from background noise, other voices, echo, and codec artifacts — reduces word error rate from 15-30% down to ~5% in noisy real-world conditions across languages and accents
• Turn Prediction v3: predicts end-of-turn from speech prosody and rhythm in under 200ms — catches 47% more true turn-shifts than v2 without extra false positives; multilingual (13+ languages); runs on CPU at 30MB with no transcription needed
• Interruption Prediction v1 (industry-first audio-only model): tells the difference between backchannels (uh-huh, yeah) and real interruptions in under 1 second with under 6% false positives — plus Signal Detectors for TTS detection, gender, and accent identification

Discussion