Merge pull request 'pitch: treat bleed in the vocal stem as silence' (#15) from pitch-silence into main

Reviewed-on: #15
Reviewed-by: branderson <brad@branderson.io>

src/songtrace/stages/pitch.py

@@ -6,6 +6,11 @@ import numpy as np
 from ..models import on_release
 from ..pipeline import Ctx, Stage
 
+# Separated vocal stems keep faint bleed that crepe still finds a pitch in, so
+# frames this far below the song's loud vocal level count as silence. Relative
+# rather than absolute, so it holds however loud the song was mastered.
+SILENCE_BELOW_LOUD_DB = 25.0
+
 
 @on_release
 def unload_model() -> None:
@@ -18,6 +23,11 @@ def unload_model() -> None:
             delattr(torchcrepe.infer, attr)
 
 
+def mute_silence(conf: np.ndarray, loudness: np.ndarray) -> None:
+    """Zero the confidence of frames quiet enough to be bleed, in place."""
+    conf[loudness < np.quantile(loudness, 0.9) - SILENCE_BELOW_LOUD_DB] = 0.0
+
+
 def track_pitch(vocals_path: Path, model: str, device: str):
     import torchaudio
     import torchcrepe
@@ -33,6 +43,7 @@ def track_pitch(vocals_path: Path, model: str, device: str):
         return_periodicity=True, batch_size=512)
     f0 = f0[0].cpu().numpy()
     conf = periodicity[0].cpu().numpy()
+    mute_silence(conf, torchcrepe.loudness.a_weighted(waveform, 16000, 160)[0].cpu().numpy())
     times = np.arange(len(f0)) * 0.01
     return times, f0, conf
 

tests/test_pitch.py

@@ -17,3 +17,21 @@ def test_pitch_stage_writes_npz(tmp_path, monkeypatch):
     data = np.load(work / "pitch.npz")
     assert np.allclose(data["f0"], 220.0)
     assert data["times"].shape == data["conf"].shape
+
+
+def test_mute_silence_drops_bleed_but_keeps_quiet_singing():
+    # 60 frames of singing around -30 dB, one quiet sung frame at -50 dB, and
+    # 40 frames of bleed at -60 dB: only the bleed is more than 25 dB below
+    # the song's loud level.
+    loudness = np.array([-30.0] * 60 + [-50.0] + [-60.0] * 40)
+    conf = np.full_like(loudness, 0.8)
+    pitch.mute_silence(conf, loudness)
+    assert (conf[:61] == 0.8).all()
+    assert (conf[61:] == 0.0).all()
+
+
+def test_mute_silence_is_relative_to_the_songs_level():
+    loudness = np.array([-45.0] * 60 + [-75.0] * 40)
+    conf = np.full_like(loudness, 0.8)
+    pitch.mute_silence(conf, loudness)
+    assert (conf[:60] == 0.8).all() and (conf[60:] == 0.0).all()