Zyphra open-sources textnorm tool, lifting TTS structured-data accuracy by up to 22 points

kastnerkyle · x · 2026-11-16

Zyphra researchers argue text normalization is underrated in TTS: on Gradium's TTS Hard Cases set, properly normalizing inputs solves most failures. They released zyphra-textnorm, which handles normalization without added latency. On a 2,500-item structured-data test set across EN/FR/ES/PT/DE, Zonos correctness rose from 45.64% to 67.60% (+21.96pp) and Qwen from 48.52% to 59.08% (+10.56pp), with all five languages improving.

Original post →

More from Multimodal

Multimodal channel →