Confidence-Modulated Speculative Decoding for Large Language Models