CueSheet Multi-genre concert segment detection · live demo

Per-second sliding window. Each second t, the audio encoder consumes the trailing 10 s of audio (window [t-9, t+1]) and emits one prediction. There is no peek into the future — only what the system has actually heard.

Cold start. At the very first second after a jump or session start, only ~1 s of real audio is available; the encoder window is left-padded with silence. Expect the first ~5-10 s of predictions to be biased toward whatever the encoder reports for mostly-silent input (often Ambient) before the trailing buffer fills.

HMM memory. The 4-class encoder posterior is smoothed by an online causal HMM forward pass (stay = 0.95). The HMM carries one floating-point state vector (alpha) across seconds — this is the only thing the live system "remembers". On a jump or mode switch, alpha is reset to a uniform prior so no memory leaks across sessions.

Heuristics. Two of the six classes — Pre-concert and Intermission — are not encoder outputs; they're temporal rules applied on top of the smoothed segment track (first sustained run of Performance ends the pre-concert lead-in; long low-activity gaps between Performance segments are intermissions).