Reproduce “Voice Conversion Improves Cross-Domain Robustness for Spoken Arabic Dialect Identification” on OpenResearch

About this paper

Badr M. Abdullah, Matthew Baas, Bernd Möbius, Dietrich Klakow

Researchers from Saarland University and Camb.AI demonstrate that voice conversion can dramatically improve Arabic Dialect Identification systems' cross-domain robustness, achieving +34.07% relative improvement in average accuracy across unseen domains (radio, theater, TEDx talks) by training models on re-synthesized speech using kNN-VC with multiple target speakers, reaching 85.32% in-domain and 80.73% cross-domain accuracy that surpasses traditional augmentation methods and sets new state-of-the-art performance on the ADI-5 benchmark, with controlled experiments revealing that voice conversion mitigates dataset speaker bias by forcing models to learn genuine dialectal features rather than exploiting speaker-specific acoustic shortcuts.

Key insights

• Voice conversion acts as a powerful bias mitigation strategy by disentangling speaker identity from dialectal cues, preventing models from learning superficial speaker-based shortcuts
• Models trained solely on re-synthesized speech with unified target speakers across dialects achieve competitive performance, while those with dialect-specific target speakers collapse to chance-level accuracy
• Cross-domain robustness requires understanding and strategically modifying training data properties rather than relying solely on complex model architectures

Summary by alphaXiv.