Background Speech Wrecks STT Accuracy. VIVA 2.5 Cuts Errors by 70%

We've just released VIVA 2.5, already running across 1B+ minutes of voice-AI conversations every month. This release includes two new models: Voice Isolation 2.5 and Voice Isolation 2.5 Lite.
Here's what they bring:

• 57% average WER reduction across all tested conditions

• STT engines become usable again. WER drops from 61% to 14.9% when competing voices are present — the condition that previously broke most STT pipelines

• No hurt on clean audio. On single-speaker recordings with clean audio, Voice Isolation 2.5 doesn't hurt accuracy. This is a significant update from Voice Isolation 2.1

• Voice Isolation 2.5 Lite delivers comparable accuracy while being 3.5x smaller, for CPU-constrained and edge deployments

Full details on the blog → https://krisp.ai/blog/voice-isolation-2-5/