Evaluating Bias in Spoken Dialogue LLMs for Real-World Decisions and Recommendations