EmotionThinker 實現可解釋語音情感 AI

🔑 增強重點摘要

• •EmotionThinker constructs EmotionCoT-35K, a 35K-scale emotional reasoning dataset featuring Chain-of-Thought annotations and detailed captions for training explainable SER[1][2][3].
• •Introduces GRPO-PTR, a novel RL algorithm that extends Group-Relative-Policy-Optimization with progressive trust-aware reasoning rewards based on multi-dimensional criteria[2][3].
• •Develops EmotionThinker-Base, a prosody-enhanced foundation model via supervised fine-tuning to better discriminate fine-grained prosodic patterns like intonation trends[1][2][6].

🛠️ 技術深入

• •Three-stage framework: (1) Prosody-enhanced pretraining/SFT on EmotionThinker-Base to improve acoustic cue perception; (2) Dataset construction with EmotionCoT-35K for CoT reasoning; (3) RL optimization using GRPO-PTR with dynamic trustworthiness-weighted rewards aligning reasoning and outcomes[1][2].
• •GRPO-PTR differs from standard GRPO by introducing reasoning rewards evaluated via a multi-dimensional reward model, preventing suboptimal strategies that match outcomes but lack trustworthy reasoning[2][3].
• •Outperforms 16 open-source SpeechLLMs on emotion accuracy and explanation quality across multiple benchmarks[1][2].
• •Leverages cues from speaker traits, prosody (e.g., intonation), semantics, and logic for generating predictions with explanatory reasoning[1][2].

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線