Reproduce “Learning Speech Attribute Representations via Multi-Teacher Distillation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Maryem Bouziane, Salima Mdhaffar, Yannick Estève
The authors present a shared speech encoder with four attribute-specific branches for semantic, speaker, language, and emotion representations, using teacher embeddings for supervision. On VoxCeleb1-O speaker verification, the model reaches 0.915% EER, versus 0.899% for the single-task speaker student. Language identification reaches 6.39% error with the combined semantic-plus-language representation; results support four evaluated tasks, not measured efficiency gains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.