TurTalker: Efficient Speech-Driven 3D Talking Face for Turkish via Rectified Flow Matching
IEEE Access, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/access.2026.3744983
- Dergi Adı: IEEE Access
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Anahtar Kelimeler: 3D talking face database for Turkish, Face animation, Multimodal interfaces, Speech-driven 3D talking face
- Boğaziçi Üniversitesi Adresli: Evet
Özet
Diffusion models have proven successful for speech-driven 3D facial animation, but their large number of denoising steps imposes a per-window latency that is incompatible with interactive use. We present TurTalker, a FLAME-based system that adopts the conditioning and decoder design of state-of-the-art diffusion-based methods, but replaces their generative backbone with few-step rectified flow matching. Progress in this field has concentrated almost exclusively on English, and production in under-resourced languages remains a challenge even with multilingual speech encoders. We therefore benchmark TurTalker on the Turkish SUDFace3D corpus, a speaker-diverse non-English testbed. Our design is motivated by two empirical observations. The first concerns the output parameterization, which decides whether few-step flow matching is competitive at all. Predicting clean motion rather than velocity improves every metric at the same step count. The second is that the audio encoder contributes more than the choice of generative backbone. Replacing mHuBERT-147 with XLS-R-300M improves validation lip accuracy by more than 22% for both backbones. The recommended few-step configuration surpasses the validation-tuned 500-step diffusion in lip accuracy across five training seeds while using ≈71× fewer network evaluations at the motion-generation stage, so that motion generator is no longer the computational bottleneck.