To overcome the limitations of existing subband-based encoders (fixed input length and lack of explicit frequency-position encoding), this paper proposes a novel baseline model, ECHO, which integrates an advanced band-segmentation architecture with relative frequency-position embedding. ECHO supports arbitrary-length inputs without padding or segmentation and generates concise embeddings that preserve temporal and spectral fidelity. We experimentally demonstrate state-of-the-art performance on anomaly detection and fault identification using the large-scale benchmark SIREN, which incorporates diverse datasets (including all DCASE task 2 challenges (2020-2025) and widely used industrial signal corpora). ECHO is available as open source.
