Gamsing

The system extracts time‑frequency features, drum onset information, and short‑term energy changes, converting raw audio into structured data that AI can interpret. A two‑stage CRNN model then performs hierarchical recognition: Stage 1 identifies silence and active drum performance regions. Stage 2 distinguishes steady rhythm sections from transitional drum fills. Background noise, decaying sounds, and low‑level interference are effectively filtered out. **Core Technologies:** Audio feature extraction, drum onset detection, short‑term energy analysis, and two‑stage CRNN classification.

Instead of cutting audio at fixed time intervals, the system converts frame‑level predictions into continuous rhythm, fill, and silence sections based on the actual musical structure. Each initial boundary is refined through energy analysis and localized searching: - Snaps cut points to natural energy valleys - Relabels and restores continuous silence regions - Merges short fragments caused by false detection - Divides overly long sections into more meaningful musical units **Core Technologies:** Temporal segmentation, candidate refinement, energy‑valley alignment, silence relabeling, fragment merging, and long‑section splitting.

After segmentation, the system analyzes the drum‑onset distribution, accent positions, note density, and timing relationships within each section. It can then: - Compare structural similarities between rhythm sections - Group similar grooves into the same rhythm pattern - Identify repeated or closely related sections - Estimate tempo and correct common half‑time or double‑time errors - Produce structured results with groove and tempo metadata **Core Technologies:** Onset‑based rhythm representation, similarity calculation, pattern clustering, tempo estimation, and reference alignment。