Ningyuan Yang

P2Flow is introduced, a phoneme-aware progressive flow matching (FM) framework designed for extreme SSR with three main strategies, which leverages phonetic information to reconstruct missing spectral components and employs a progressive architectural design that hierarchically restores distinct frequency regions.

Current LALMs can identify basic audio content but often fail to diagnose and reason about degradation, revealing a significant blind spot in their capabilities.

Forget manual curation—aligning policy gradients with a validation set adaptively selects RL training data, leading to more stable LLM training and improved performance.