Can AI understand not just *what* you say, but *how* you say it? New research from Meta explores the fascinating ability of "frozen" large language models (LLMs) to perceive the emotional nuances in human speech. Traditionally, interacting with an LLM has been like communicating through text messages – the AI understands the words, but misses the tone. However, this new research suggests that even without retraining the core model, LLMs can be adapted to recognize emotions and speaking styles in spoken prompts. The researchers achieved this by using a separate speech encoder, trained to translate audio into special tokens that carry both semantic and paralinguistic information. These tokens are then fed to the frozen LLM, allowing it to interpret the emotional context. Think of it as giving the LLM "emotional hearing." By aligning the LLM's responses to expressive speech with those to text prompts containing explicit emotional tags (like "" or ""), the researchers effectively taught the encoder to capture and convey these emotional cues. Impressively, this approach not only allows the LLM to respond more empathetically and appropriately to the tone of the speech, but also improves its performance on related tasks like speech emotion recognition. This breakthrough has significant implications for the future of AI interactions. Imagine voice assistants that understand your frustration when you're struggling with a problem, or chatbots that can offer comforting words when you're feeling down. This research suggests we're one step closer to a future where AI truly understands us, not just our words, but our emotions too.