AHAMask: Reliable Task Specification for Large Audio Language Models without Instructions