Synthetic Speech, Real Signal: Paralinguistic Preservation and Cross-Lingual Augmentation via Voice Cloning
Merged summary
TL;DR - This study finds that most tested voice-cloning models preserve clinically relevant paralinguistic signals with modest degradation. Cloned English-to-Japanese speech also improves depression and anxiety detection over direct cross-lingual transfer.
- Benchmarks eight voice-cloning models across five public and clinical paralinguistic tasks.
- Evaluates downstream task performance rather than only intelligibility or speaker similarity.
- Demonstrates potential for augmenting scarce clinical speech data in low-resource languages.
Sources (1)
Synthetic Speech, Real Signal: Paralinguistic Preservation and Cross-Lingual Augmentation via Voice Cloning
TL;DR - This study finds that most tested voice-cloning models preserve clinically relevant paralinguistic signals with modest degradation. Cloned English-to-Japanese speech also improves depression and anxiety detection over direct cross-lingual transfer.
- Benchmarks eight voice-cloning models across five public and clinical paralinguistic tasks.
- Evaluates downstream task performance rather than only intelligibility or speaker similarity.
- Demonstrates potential for augmenting scarce clinical speech data in low-resource languages.