Study finds speech domain explains 4x more variance than model choice in TTS cloning
kastnerkyle · x · 2026-10-19
A new arXiv paper systematically evaluates Qwen3-TTS's two voice-cloning interfaces against Chatterbox, XTTS-v2 and F5-TTS.
- ICL vs X-Vector: neither dominates — ICL clones timbre better with correct transcripts, X-Vector clones timing better; ICL degrades sharply on wrong short transcripts
- Domain matters most: speech domain (read/meeting/emotional/ATC) explains 4x more variance than system choice
- Qwen3-TTS leads on intelligibility and pitch but trails on prosody; objective metrics rank systems differently from listeners
- Safety: an off-the-shelf countermeasure flags its clones no more than real recordings; ATC radio fails for every system
Code and audio samples are released.
More from Multimodal
- Vintage Photocopy prompt turns any subject into distressed old-copier art — LudovicCreator · 2026-10-19
- Same-prompt test: Kling 4.0 vs Seedance 2.0 vs MiniMax H3 on fashion editorial — FellMentKE · 2026-10-19
- Short film storyWeave: prompts, dialogue and VFX largely generated by Opus — hq4ai · 2026-10-19
- Ara 3D Gen open-sourced: generate rigged, animatable realistic 3D humans locally — Fowe · 2026-10-19
- Gothic AI illustrations: a dead-letter postman and a forgotten prima ballerina — azed_ai · 2026-10-19
- Opus 5.5 edits video in Claude Code: one prompt, 15 jobs, 100/100 QA score — socialwithaayan · 2026-10-19