UltraVoice Dataset for Fine-Grained Speech Style Control
September 1, 2026
UltraVoice provides 830 hours of speech dialogues covering six stylistic dimensions: emotion, speed, volume, accent, language, and composite styles. Fine-tuning SLAM-Omni or VocalNet on this dataset improves Mean Opinion Score (MOS) by up to 42.33% without losing conversational reasoning.
HOW THIS AFFECTS YOU
●
builderYou can use this dataset to add nuanced prosody and emotional control to your spoken dialogue agents.
●
designerThis enables more human-like and expressive vocal personas in user interfaces.