Imagine a world where AI can seamlessly understand and translate spoken language, transforming conversations in real-time. This future hinges on effectively bridging speech and text, a challenge researchers are tackling by connecting Speech Foundation Models (SFMs) with Large Language Models (LLMs). SFMs excel at turning audio into a format AI can interpret, while LLMs are masters of language generation and understanding. Connecting them seems like a natural fit, but the 'how' has been a key question. Researchers have been experimenting with 'adapter' modules to link SFMs and LLMs, but until now, the impact of design choices on overall performance remained a mystery. A new study delves into this question, comparing different combinations of SFMs, LLMs, and adapter designs for speech recognition and translation tasks. Surprisingly, the type of SFM played a bigger role than expected, overshadowing the choice of LLM or adapter in overall performance. SeamlessM4T consistently outperformed Whisper, another popular SFM. While adapter selection mattered, no single best option emerged. Interestingly, the study found that reducing sequence length mismatch between speech and text is less important than previously thought, which opens new avenues for more efficient designs. These findings highlight the complex interplay between model components in the quest for more robust and fluent speech AI. Future research will likely focus on optimizing the SFM-LLM connection to unlock the full potential of these combined powerful models, paving the way for smoother, more natural human-computer interactions.