Expand description
Multi-voice pitch-contour tracking (polyphonic extension of Melodia).
Melodia’s mono predominant-melody stage collapses each frame to a single F0
and threads it through time. Polyphonic material has several concurrent
lines, so this tracker takes the per-frame multi-pitch estimates (e.g.
from super::multipitch_klapuri, one row of F0s per analysis frame) and
threads them into up to n_voices continuous pitch tracks by nearest-pitch
continuity — the many-voices analogue of contour tracking.
EPISTEMIC RULE (declared): the tracks are proposals. The max-polyphony
assumption is explicit in n_voices: at most that many concurrent lines
are posited, and when a frame offers more estimates than free voices the
surplus is dropped rather than forced onto an unrelated voice. A voice with
no matching estimate in a frame abstains for that frame (0.0), and is
freed after MAX_GAP silent frames so a genuinely new line can reuse the
slot. Zero input estimates → zero tracks.
Zero-heap hot path: all per-voice / per-estimate state lives in fixed stack
arrays bounded by MAX_VOICES / MAX_ESTIMATES; the caller owns the
frame_f0 input and the out_tracks output. Nothing is allocated.
Constants§
- MAX_
ESTIMATES - Upper bound on per-frame F0 estimates (stack-array sizing).
- MAX_GAP
- Consecutive unmatched frames after which a voice is freed for reuse.
- MAX_
VOICES - Upper bound on concurrent voices (stack-array sizing).
Functions§
- track_
multi_ pitch - Track multiple concurrent pitch contours across frames.