Root-mean-square amplitude of the window, normalized to ~[0, 1].
Peak absolute amplitude of the window, normalized to ~[0, 1].
Share of amplitude-envelope modulation energy in the 3–8 Hz syllable band versus the 0.5–2.5 Hz note band. High for speech, low for sustained singing.
Share of 10 ms frames quieter than half the window mean (speech pauses).
Share of pitch frames judged voiced (periodic). Singing is near-continuous.
Share of pitch frames belonging to a sustained note: a run held within a narrow pitch band for at least the configured minimum duration.
Mean distance of voiced pitches from the nearest semitone of the best-fitting tuning grid, in cents (0–25; lower means more musically tuned).
Number of pitch frames analysed in this window.
Per-window acoustic features used to separate speech from singing.
All ratios are in
[0, 1]unless noted. Values are computed over one analysis window (typically 2 s of 16 kHz mono audio).