Reproduce “Flow Matching-Based Speech Source Separation with Best-of-N Biometric Sampling” 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
Anastasia Zorkina, Alexandr Anikin, Nikita Khmelev, Anastasiya Korenevskaya, Sergey Novoselov, Vladimir Volokhov, Maxim Korenevsky, Yuriy Matveev
Researchers at ITMO University and Speech Technology Center Ltd. developed a flow matching-based speech source separation system that integrates biometric sampling and chunk alignment to produce consistently ordered outputs from generative models. The system lowered ASR character/phrase error rates to 3.84% and speaker verification equal error rates to 0.39% on the Libri2Mix clean test set.
Key insights
- Biometric source ordering during training, utilizing a pre-trained speaker encoder, effectively imposes a canonical target order, addressing permutation ambiguity.
- Selecting the separation candidate that maximizes speaker disentanglement (least similar speaker embeddings) effectively mitigates the stochasticity of generative models without requiring ground truth.
- Integrating chunk-wise processing with a speaker-recognition-based alignment mechanism allows robust and consistent source separation for long-form audio, crucial for real-world deployment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.