A podcast discussing a new approach to audio-driven portrait animation called Sonic, focusing on global audio perception rather than visual cues. The discussion covers the methodology behind Sonic, including context-enhanced audio learning, motion-decoupled controllers, and time-aware position shift fusion, as well as experimental results.

A podcast discussing Titans: Learning to Memorize at Test Time.

A podcast discussing Tensor Product Attention, a novel attention mechanism for Large Language Models.

A podcast discussing Classifier-Free Diffusion Guidance, a method for improving sample quality in diffusion models without relying on a separate classifier.

本播客深入探讨了阿里巴巴 Tongyi Lab 的 MinMo 模型，这是一种旨在实现无缝语音交互的多模态大型语言模型。我们讨论了其架构、训练过程以及在各种语音任务中的性能，包括语音识别、翻译、情感识别和全双工对话。

A podcast discussing the paper Titans: Learning to Memorize at Test Time.

探讨iSTFTNet如何通过逆短时傅里叶变换优化梅尔频谱声码器，提高速度和效率。

深入探讨微软最新发布的Phi-4大型语言模型，了解其在数据质量、合成数据、训练方法和后训练优化方面的创新。 我们将分析其在各种基准测试上的表现，以及它如何超越以往的Phi系列模型，甚至在某些领域超越更大的模型。

A podcast discussion about the StyleTTS 2 model for text-to-speech synthesis, focusing on its innovative use of style diffusion and adversarial training with large speech language models to achieve human-level performance.

本播客深入探讨了语音对话模型的最新进展，包括其功能、表示形式、训练范式以及流媒体和交互能力。