Imagine talking to your virtual assistant in a bustling cafe. It’s tough enough for *you* to filter out the surrounding chatter, but how does your device know you're talking to *it* and not the barista? This is the challenge of device-directed speech detection (DDSD), particularly in follow-up conversations where you don’t repeat the wake word. New research from Apple explores innovative ways to use Large Language Models (LLMs) to make DDSD more accurate in these challenging scenarios. Instead of just processing individual phrases, they're teaching LLMs to understand the context of the whole conversation. For example, if you ask Siri to “play music” and then say “turn it up,” the LLM uses the first query to understand that the follow-up is also directed at the device, even without hearing “Hey Siri” again. The researchers also address the imperfections of automatic speech recognition (ASR). ASR systems, which convert speech to text, sometimes make mistakes. To overcome this, they provide the LLM with multiple possible transcriptions (called n-best hypotheses), along with their probabilities. This helps the LLM consider alternatives and choose the most likely interpretation. The Apple team experimented with different ways of training LLMs, including both direct prompting (giving the LLM instructions) and fine-tuning a classifier on top of the LLM. They found significant improvements, achieving a 20-40% reduction in errors when using context and n-best hypotheses together. This breakthrough could mean fewer frustrating misinterpretations and a smoother experience when using voice assistants in everyday life. While the focus here is on pairs of queries, the future looks even brighter. Integrating additional signals like the assistant's responses, acoustic features, and speaker identification promises even more accurate DDSD in the future. This is a big step towards making conversations with our devices feel more natural and human-like.