This paper explores Reservoir Computing as a feature-free method for raw audio signal classification. It shows that parallel deep reservoir architectures outperform shallow and sequential ones in accuracy while maintaining low complexity, enabling efficient, low-power audio processing with minimal preprocessing.