What speech languages and emotions was Audio2Face-3D trained on, and how does emotion control work?

Problem. Audio-driven faces inherit the training data's languages and emotional range. Conditions. Choosing a model for a non-English or highly emotional character. Method. The paper's capture used professional actors speaking English or Mandarin sentences (about 50 to 70 per actor, 3 to 15 s) across 11 emotional states. Emotion enters as a label vector that can be keyframed or blended; Audio2Emotion can detect it from audio, but its classifier uses six classes and training corpora restricted to European languages. The tongue was not captured: an animator animated it from the tracked jaw. Failure modes. Other languages may show lip shapes outside the training distribution (inference from the stated training languages); emotion labels outside the 11 states are not trained. Limitations. The paper does not report per-language error, so language bias is inferred from the data description. Version/date. arXiv 2508.16401; model v3.0. Curator qualifications: Some cited evidence could not be re-located by curator GET (see mechanical_checks); no video sources. Second-tier recheck (C3R-0614a) re-located the excerpts by timed GET. Limitations: Language bias is my inference; the authors give no per-language results. The open v3.0 weights may use different data than the paper's captures. Evidence status: statements located in the cited sources by a correlated desk review (Claude); not executed; not independently reviewed.

{
  "citations": [
    {
      "url": "https://arxiv.org/pdf/2508.16401",
      "note": "Sec. 3.1 Data Capture; Sec. 3.1 Data Capture; Sec. 3.1 Tongue; Sec. 7 Results; Appendix A.3; researcher reasoning"
    }
  ]
}