Reproduce “Profile-Error-Tolerant Target-Speaker Voice Activity Detection” 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
Dongmei Wang, Xiong Xiao, Naoyuki Kanda, Midia Yousefi, Takuya Yoshioka, Jian Wu
This paper presents Profile-Error-Tolerant Target-Speaker Voice Activity Detection (PET-TSVAD), an extension designed to enhance the robustness of TS-VAD systems against imperfect speaker profiles generated by initial diarization steps. The framework consistently reduced the Diarization Error Rate (DER) on VoxConverse and DIHARD-I datasets, demonstrating improved handling of speaker confusion and false alarm errors.
Key insights
- The introduction of pseudo-speaker profiles effectively provides a mechanism for the TS-VAD system to account for and detect speakers who were entirely missed during the first-pass diarization.
- Training with a diversified set of error-containing speaker profiles significantly improves the model's robustness and generalization capabilities when facing imperfect input profiles in real-world scenarios.
- Permutation Invariant Training is crucial for enabling the model to learn effectively under conditions where the correspondence between input profiles and true speakers is ambiguous or flawed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.