mroctopus / transcriber /optimize.py
Ewan
Winner-takes-all drum transcription + harsher piano rhythm grid
8b89f6d
"""Optimize MIDI transcription by correcting onsets, cleaning artifacts, and
ensuring rhythmic accuracy against the original audio."""
import copy
from pathlib import Path
import numpy as np
import pretty_midi
import librosa
from collections import Counter
def remove_leading_silence_notes(midi_data, y, sr):
"""Remove notes that appear during silence/noise before the music starts.
Finds the first moment of real musical energy and removes any MIDI notes
before that point (typically microphone rumble / low-freq noise artifacts).
Always preserves the first detected MIDI note to prevent eating the opening.
"""
midi_out = copy.deepcopy(midi_data)
# Compute RMS in 50ms windows
hop = int(0.05 * sr)
rms = np.array([
np.sqrt(np.mean(y[i * hop:(i + 1) * hop] ** 2))
for i in range(len(y) // hop)
])
if len(rms) == 0:
return midi_out, 0, 0.0
# Music starts when RMS first exceeds 5% of the peak energy
# (reduced from 10% to avoid eating quiet openings)
max_rms = np.max(rms)
music_start = 0.0
for i, r in enumerate(rms):
if r > max_rms * 0.05:
music_start = i * 0.05
break
if music_start < 0.1:
return midi_out, 0, music_start
# Find the earliest MIDI note onset — always protect it
all_notes = sorted(
[n for inst in midi_out.instruments for n in inst.notes],
key=lambda n: n.start
)
earliest_onset = all_notes[0].start if all_notes else 0.0
# If the "silence" region would eat the first note, clamp music_start
if music_start > earliest_onset:
music_start = earliest_onset
if music_start < 0.1:
return midi_out, 0, music_start
removed = 0
for instrument in midi_out.instruments:
filtered = []
for note in instrument.notes:
if note.start < music_start:
removed += 1
else:
filtered.append(note)
instrument.notes = filtered
return midi_out, removed, music_start
def remove_trailing_silence_notes(midi_data, y, sr):
"""Remove notes that appear during the audio fade-out/silence at the end.
Uses a 2% RMS threshold (reduced from 5%) and adds a 3-second protection
zone after the detected music end to preserve natural piano decay/sustain.
"""
midi_out = copy.deepcopy(midi_data)
hop = int(0.05 * sr)
rms = np.array([
np.sqrt(np.mean(y[i * hop:(i + 1) * hop] ** 2))
for i in range(len(y) // hop)
])
if len(rms) == 0:
return midi_out, 0, len(y) / sr
max_rms = np.max(rms)
# Find the last moment where RMS exceeds 2% of peak (searching backwards)
# Reduced from 5% to preserve quiet endings, fade-outs, and final sustain
music_end = len(y) / sr
for i in range(len(rms) - 1, -1, -1):
if rms[i] > max_rms * 0.02:
# Add 3-second protection zone for natural piano decay
music_end = (i + 1) * 0.05 + 3.0
break
# Clamp to actual audio duration
music_end = min(music_end, len(y) / sr)
removed = 0
for instrument in midi_out.instruments:
filtered = []
for note in instrument.notes:
if note.start > music_end:
removed += 1
else:
filtered.append(note)
instrument.notes = filtered
return midi_out, removed, music_end
def remove_low_energy_notes(midi_data, y, sr, hop_length=512):
"""Remove notes whose onsets don't correspond to real audio energy.
Computes the onset strength envelope and removes notes at times
where the audio shows no significant onset energy. This catches
basic-pitch hallucinations that appear at normal pitches but have
no corresponding audio event.
Uses an adaptive threshold based on the recording's onset strength
distribution (15th percentile), so it works equally well on loud
and quiet recordings.
"""
midi_out = copy.deepcopy(midi_data)
onset_env = librosa.onset.onset_strength(y=y, sr=sr, hop_length=hop_length)
onset_times = librosa.frames_to_time(
np.arange(len(onset_env)), sr=sr, hop_length=hop_length
)
removed = 0
for instrument in midi_out.instruments:
# First pass: measure strength per note
note_strengths = []
for note in instrument.notes:
frame = np.argmin(np.abs(onset_times - note.start))
lo = max(0, frame - 2)
hi = min(len(onset_env), frame + 3)
strength = float(np.max(onset_env[lo:hi]))
note_strengths.append(strength)
if not note_strengths:
continue
# Adaptive threshold: 15th percentile of note onset strengths
# This adapts to the recording's volume — quiet recordings get
# a lower threshold, loud recordings get a higher one.
# Floor at 0.5 to always catch clearly silent hallucinations.
strength_threshold = max(0.5, float(np.percentile(note_strengths, 15)))
filtered = []
for idx, note in enumerate(instrument.notes):
if note_strengths[idx] >= strength_threshold:
filtered.append(note)
else:
# Keep notes that are part of a chord with a strong onset
chord_has_energy = False
for other_idx, other in enumerate(instrument.notes):
if other is note:
continue
if abs(other.start - note.start) < 0.03 and note_strengths[other_idx] >= strength_threshold:
chord_has_energy = True
break
if chord_has_energy:
filtered.append(note)
else:
removed += 1
instrument.notes = filtered
return midi_out, removed
def remove_harmonic_ghosts(midi_data, y=None, sr=22050, hop_length=512):
"""Remove notes that are harmonic doublings of louder lower notes.
Two-stage detector:
1. Pairwise: for notes at harmonic intervals (7, 12, 19, 24 semitones),
remove the upper note if it's clearly a harmonic ghost.
2. Spectral masking: when bass + melody overlap (two-hand texture),
check if upper notes can be explained by the harmonic series of
strong lower notes. This catches ghost notes that the pairwise
detector misses because they're at non-standard intervals.
Uses CQT energy to protect strong notes: if the CQT shows the note
has strong independent energy distinct from what the lower note's
harmonics would produce, it's a real played note.
"""
midi_out = copy.deepcopy(midi_data)
removed = 0
harmonic_intervals = {7, 12, 19, 24}
# Compute CQT for energy verification if audio provided
C_db = None
N_BINS = 0
if y is not None:
N_BINS = 88 * 3
FMIN = librosa.note_to_hz('A0')
C = np.abs(librosa.cqt(
y, sr=sr, hop_length=hop_length,
fmin=FMIN, n_bins=N_BINS, bins_per_octave=36,
))
C_db = librosa.amplitude_to_db(C, ref=np.max(C))
for instrument in midi_out.instruments:
notes = sorted(instrument.notes, key=lambda n: n.start)
to_remove = set()
for i, note in enumerate(notes):
if i in to_remove:
continue
if note.pitch < 48:
continue
# Check CQT energy — protect notes with moderate+ energy
if C_db is not None:
fund_bin = (note.pitch - 21) * 3 + 1
if 0 <= fund_bin < C_db.shape[0]:
start_frame = max(0, int(note.start * sr / hop_length))
end_frame = min(C_db.shape[1], start_frame + max(1, int(0.2 * sr / hop_length)))
lo = max(0, fund_bin - 1)
hi = min(C_db.shape[0], fund_bin + 2)
onset_db = float(np.max(C_db[lo:hi, start_frame:end_frame]))
if onset_db > -12.0:
# Real CQT energy present — keep this note
continue
for j, other in enumerate(notes):
if i == j or j in to_remove:
continue
if abs(other.start - note.start) > 0.10:
continue
diff = note.pitch - other.pitch
if diff in harmonic_intervals and diff > 0:
ratio = note.velocity / max(1, other.velocity)
if note.pitch >= 72:
# C5+: only remove if much quieter than the lower note
if ratio < 0.55:
to_remove.add(i)
break
elif other.pitch < 48:
# Sub-bass pairs: keep tighter — sub-bass ghosts are common
if ratio < 0.85:
to_remove.add(i)
break
else:
# General: only remove if clearly quieter
if ratio < 0.55:
to_remove.add(i)
break
# Stage 2: Spectral masking for two-hand texture
# When bass (< MIDI 55) and melody (>= MIDI 60) overlap, bass harmonics
# can produce ghost notes in the melody range. Check if a mid-range note
# is explainable as a harmonic partial of a concurrent bass note.
if C_db is not None:
remaining = [(k, n) for k, n in enumerate(notes) if k not in to_remove]
bass_notes = [(k, n) for k, n in remaining if n.pitch < 55]
mid_notes = [(k, n) for k, n in remaining if 55 <= n.pitch < 72]
for mid_k, mid_n in mid_notes:
if mid_k in to_remove:
continue
for bass_k, bass_n in bass_notes:
if abs(bass_n.start - mid_n.start) > 0.05:
continue
# Check if mid_n.pitch matches any harmonic partial of bass_n
# Harmonics: 2nd (+12), 3rd (+19), 4th (+24), 5th (+28), 6th (+31)
bass_pitch = bass_n.pitch
harmonic_pitches = {
bass_pitch + 12, # 2nd harmonic (octave)
bass_pitch + 19, # 3rd (octave + fifth)
bass_pitch + 24, # 4th (2 octaves)
bass_pitch + 28, # 5th (2 oct + major 3rd)
bass_pitch + 31, # 6th (2 oct + fifth)
}
if mid_n.pitch in harmonic_pitches:
# This mid note matches a bass harmonic — check if
# it has independent CQT energy above the harmonic level
mid_bin = (mid_n.pitch - 21) * 3 + 1
bass_bin = (bass_pitch - 21) * 3 + 1
if 0 <= mid_bin < N_BINS and 0 <= bass_bin < N_BINS:
sf = max(0, int(mid_n.start * sr / hop_length))
ef = min(C_db.shape[1], sf + max(1, int(0.15 * sr / hop_length)))
mid_energy = float(np.max(C_db[max(0, mid_bin-1):min(N_BINS, mid_bin+2), sf:ef]))
bass_energy = float(np.max(C_db[max(0, bass_bin-1):min(N_BINS, bass_bin+2), sf:ef]))
# If bass is much louder (>8dB) and mid note is quiet,
# it's likely a harmonic ghost
if bass_energy - mid_energy > 8.0 and mid_n.velocity < bass_n.velocity * 0.7:
to_remove.add(mid_k)
break
instrument.notes = [n for k, n in enumerate(notes) if k not in to_remove]
removed += len(to_remove)
return midi_out, removed
def remove_phantom_notes(midi_data, max_pitch=None):
"""Remove high-register notes that are likely harmonic artifacts.
Uses multiple factors to distinguish real notes from phantoms:
- Must be above the 95th percentile pitch
- Must be rare (< 3 occurrences at that exact pitch)
- Must have low velocity (< 40)
- Must be isolated (no other notes within 2 semitones and 500ms)
"""
midi_out = copy.deepcopy(midi_data)
all_notes = [(n, i) for i, inst in enumerate(midi_out.instruments) for n in inst.notes]
all_pitches = [n.pitch for n, _ in all_notes]
if not all_pitches:
return midi_out, 0
if max_pitch is None:
max_pitch = int(np.percentile(all_pitches, 95))
pitch_counts = Counter(all_pitches)
# Build a time-sorted list for neighbor checking
time_sorted = sorted(all_notes, key=lambda x: x[0].start)
def is_isolated(note, all_sorted):
"""Check if a note has no other notes nearby (within 100ms).
A note in a chord or musical event is not isolated, regardless
of the pitch of its neighbors. This prevents falsely removing
high notes that are part of chords with lower-pitched notes.
"""
for other, _ in all_sorted:
if other is note:
continue
if other.start > note.start + 0.1:
break
if abs(other.start - note.start) < 0.1:
return False
return True
removed = 0
for instrument in midi_out.instruments:
filtered = []
for note in instrument.notes:
if note.pitch > max_pitch:
count = pitch_counts[note.pitch]
duration = note.end - note.start
# Higher velocity threshold for very high notes (above MIDI 80)
vel_thresh = 55 if note.pitch > 80 else 40
# Only remove if MULTIPLE indicators suggest it's a phantom:
# Very rare AND (low velocity OR very short OR isolated)
if count < 3 and (note.velocity < vel_thresh or duration < 0.08 or
is_isolated(note, time_sorted)):
removed += 1
continue
filtered.append(note)
instrument.notes = filtered
return midi_out, removed
def remove_spurious_onsets(midi_data, y, sr, ref_onsets, hop_length=512, complexity='simple'):
"""Remove MIDI notes that form false-positive onsets not backed by audio.
Analysis shows 37 extra MIDI onsets cause the biggest F1 drag (precision=0.918).
This filter targets three categories of false positives:
1. Chord fragments: notes that basic-pitch split from a real chord, creating
a separate onset within 60ms of a matched onset. These should have been
grouped with the chord.
2. Isolated ghost onsets: single-note, low-strength onsets far from any
audio onset -- pure hallucinations.
3. Short+quiet artifacts: onsets where every note is both short (<200ms)
and quiet (velocity < 50).
For complex pieces, thresholds are relaxed to preserve legitimate dense
textures that might otherwise be classified as spurious.
The filter first identifies which MIDI onsets already match audio onsets,
then only removes unmatched onsets meeting the above criteria.
"""
midi_out = copy.deepcopy(midi_data)
tolerance = 0.05
# Complexity-adjusted thresholds: complex pieces are more permissive
# to preserve legitimate dense textures
if complexity == 'complex':
strength_scale = 1.5 # require stronger evidence to remove
dist_scale = 1.4 # require further from audio onset to remove
elif complexity == 'moderate':
strength_scale = 1.2
dist_scale = 1.2
else:
strength_scale = 1.0
dist_scale = 1.0
onset_env = librosa.onset.onset_strength(y=y, sr=sr, hop_length=hop_length)
onset_times = librosa.frames_to_time(
np.arange(len(onset_env)), sr=sr, hop_length=hop_length
)
# Collect all notes and compute unique onsets
all_notes = sorted(
[n for inst in midi_out.instruments for n in inst.notes],
key=lambda n: n.start
)
midi_onsets = sorted(set(round(n.start, 4) for n in all_notes))
midi_onsets_arr = np.array(midi_onsets)
# Identify which MIDI onsets are already matched to audio onsets
matched_est = set()
for r in ref_onsets:
diffs = np.abs(midi_onsets_arr - r)
best = np.argmin(diffs)
if diffs[best] <= tolerance and best not in matched_est:
matched_est.add(best)
# For each unmatched onset, check removal criteria
onsets_to_remove = set()
for j, mo in enumerate(midi_onsets_arr):
if j in matched_est:
continue
# Get notes at this onset
onset_notes = [n for n in all_notes if abs(n.start - mo) < 0.03]
if not onset_notes:
continue
# Compute onset strength at this time
frame = np.argmin(np.abs(onset_times - mo))
lo = max(0, frame - 2)
hi = min(len(onset_env), frame + 3)
strength = float(np.max(onset_env[lo:hi]))
# Distance to nearest audio onset
diffs_audio = np.abs(ref_onsets - mo)
nearest_audio_ms = float(np.min(diffs_audio)) * 1000
# Check if near a matched MIDI onset (chord fragment)
near_matched = any(
abs(midi_onsets_arr[k] - mo) < 0.060
for k in matched_est
)
# Category 1: Chord fragment -- near a matched onset, but only if
# the onset has weak audio energy. Strong onsets near chords may be
# real grace notes or arpeggios.
if near_matched and strength < 2.0 * strength_scale:
onsets_to_remove.add(j)
continue
# Category 2: Isolated ghost -- single note, low strength or far from audio
if len(onset_notes) == 1 and (strength < 1.5 * strength_scale or nearest_audio_ms > 100 * dist_scale):
onsets_to_remove.add(j)
continue
# Category 3: Short+quiet artifact
if all(n.end - n.start < 0.2 and n.velocity < 50 for n in onset_notes):
onsets_to_remove.add(j)
continue
# Category 4: Low-velocity bass ghost -- single bass note (< MIDI 40),
# low velocity (< 35), far from audio onset. These are rumble artifacts
# that survive the energy filter.
if (len(onset_notes) == 1 and onset_notes[0].pitch < 40
and onset_notes[0].velocity < 35 and nearest_audio_ms > 60 * dist_scale):
onsets_to_remove.add(j)
continue
# Category 5: Multi-note onset far from any audio onset (> 120ms)
# with weak-to-moderate onset strength. These are chord-split artifacts
# or hallucinated events with no audio support.
if nearest_audio_ms > 120 * dist_scale and strength < 3.0 * strength_scale:
onsets_to_remove.add(j)
continue
# Category 6: All notes at this onset are very short (<50ms) —
# splinter artifacts from chord splitting, regardless of velocity.
if all(n.end - n.start < 0.05 for n in onset_notes):
onsets_to_remove.add(j)
continue
# Category 7: Moderate distance from audio (> 70ms) with weak
# onset strength — catches near-miss hallucinations that are
# just outside the 50ms matching window.
if nearest_audio_ms > 70 * dist_scale and strength < 2.5 * strength_scale:
onsets_to_remove.add(j)
continue
# Remove notes belonging to spurious onsets
times_to_remove = set(midi_onsets_arr[j] for j in onsets_to_remove)
removed = 0
for instrument in midi_out.instruments:
filtered = []
for note in instrument.notes:
note_onset = round(note.start, 4)
if any(abs(note_onset - t) < 0.03 for t in times_to_remove):
removed += 1
else:
filtered.append(note)
instrument.notes = filtered
return midi_out, removed, len(onsets_to_remove)
def remove_pitch_unconfirmed_notes(midi_data, y, sr, hop_length=512):
"""Remove notes where the CQT has no energy at their fundamental pitch.
Checks the onset region (first 200ms) of each note for CQT energy,
not the full duration. This prevents CQT-extended notes from being
falsely removed due to low average energy over their extended tail.
Targets two ranges where hallucinations concentrate:
- Sub-bass (< MIDI 40): rumble artifacts
- Upper register (> MIDI 72): harmonic doublings
Core piano range (MIDI 40-72 / E2-C5) is reliable from basic-pitch.
"""
midi_out = copy.deepcopy(midi_data)
N_BINS = 88 * 3
FMIN = librosa.note_to_hz('A0')
C = np.abs(librosa.cqt(
y, sr=sr, hop_length=hop_length,
fmin=FMIN, n_bins=N_BINS, bins_per_octave=36,
))
C_db = librosa.amplitude_to_db(C, ref=np.max(C))
# Collect all notes for chord checking
all_notes = sorted(
[n for inst in midi_out.instruments for n in inst.notes],
key=lambda n: n.start
)
# Onset region: check max energy in first 200ms
onset_frames = max(1, int(0.2 * sr / hop_length))
removed = 0
for instrument in midi_out.instruments:
filtered = []
for note in instrument.notes:
# Core mid-range (C3-C5) is reliable from basic-pitch — skip
# Bass (MIDI 40-47) gets a lenient CQT check to catch rumble
# Upper register (>72) gets checked for harmonic ghosts
if 48 <= note.pitch <= 72:
filtered.append(note)
continue
fund_bin = (note.pitch - 21) * 3 + 1
if fund_bin < 0 or fund_bin >= N_BINS:
filtered.append(note)
continue
start_frame = max(0, int(note.start * sr / hop_length))
check_end = min(C.shape[1], start_frame + onset_frames)
if start_frame >= check_end:
filtered.append(note)
continue
lo = max(0, fund_bin - 1)
hi = min(N_BINS, fund_bin + 2)
# Use max energy in onset region, not average over full duration
onset_db = float(np.max(C_db[lo:hi, start_frame:check_end]))
if note.pitch < 40:
thresh = -42.0
elif note.pitch < 48:
# Bass (C2-B2): moderate check — real bass notes have clear
# CQT energy, but threshold is lenient to keep genuine notes
thresh = -35.0
else: # > 72, upper register
thresh = -25.0
if onset_db < thresh:
# Remove if weak CQT evidence regardless of context
# Very weak = always remove; moderate weak = check isolation
if onset_db < thresh - 10:
# Extremely weak: always remove
removed += 1
continue
concurrent = sum(1 for o in all_notes
if abs(o.start - note.start) < 0.05 and o is not note)
if concurrent <= 3 or note.velocity < 55:
removed += 1
else:
filtered.append(note)
else:
filtered.append(note)
instrument.notes = filtered
return midi_out, removed
def apply_pitch_ceiling(midi_data, max_pitch=96):
"""Remove notes above a hard pitch ceiling (C7 / MIDI 96).
Only truly extreme high notes are blanket-removed. Notes between C6-C7
are kept and handled by the CQT energy filter instead, since some
(like C6, D6) can be legitimate played notes.
"""
midi_out = copy.deepcopy(midi_data)
removed = 0
for instrument in midi_out.instruments:
filtered = []
for note in instrument.notes:
if note.pitch >= max_pitch:
removed += 1
else:
filtered.append(note)
instrument.notes = filtered
return midi_out, removed
def limit_concurrent_notes(midi_data, max_per_hand=4, hand_split=60, max_left_hand=None):
"""Limit notes per chord to max_per_hand per hand.
Groups notes by onset time (within 30ms) and splits into left/right hand.
Removes excess notes — protects melody (highest RH pitch) and bass
(lowest LH pitch), then removes lowest velocity.
Args:
max_per_hand: Max notes for right hand (default 4)
max_left_hand: Max notes for left hand (defaults to max_per_hand)
"""
if max_left_hand is None:
max_left_hand = max_per_hand
midi_out = copy.deepcopy(midi_data)
removed = 0
for instrument in midi_out.instruments:
notes = sorted(instrument.notes, key=lambda n: n.start)
if not notes:
continue
chords = []
current_chord = [0]
for i in range(1, len(notes)):
if notes[i].start - notes[current_chord[0]].start < 0.03:
current_chord.append(i)
else:
chords.append(current_chord)
current_chord = [i]
chords.append(current_chord)
to_remove = set()
for chord_indices in chords:
left = [idx for idx in chord_indices if notes[idx].pitch < hand_split]
right = [idx for idx in chord_indices if notes[idx].pitch >= hand_split]
for is_right, hand_indices in [(True, right), (False, left)]:
limit = max_per_hand if is_right else max_left_hand
if len(hand_indices) <= limit:
continue
# Both hands: protect the melody (highest note)
# LH melody voice is the top line; RH melody is the top line
protected = max(hand_indices, key=lambda idx: notes[idx].pitch)
trimmable = [idx for idx in hand_indices if idx != protected]
scored = [(notes[idx].velocity, idx) for idx in trimmable]
scored.sort()
excess = len(hand_indices) - limit
for _, idx in scored[:excess]:
to_remove.add(idx)
instrument.notes = [n for k, n in enumerate(notes) if k not in to_remove]
removed += len(to_remove)
return midi_out, removed
def limit_total_concurrent(midi_data, max_per_hand=4, hand_split=60, max_left_hand=None):
"""Limit concurrent sounding notes to max_per_hand per hand.
Splits notes into left hand (< hand_split) and right hand (>= hand_split).
At each note onset, count concurrent notes in that hand. If > limit,
trim sustained notes — protect the melody (highest pitch in both hands).
Among the rest, trim lowest velocity first.
Args:
max_per_hand: Max concurrent notes for right hand (default 4)
max_left_hand: Max concurrent notes for left hand (defaults to max_per_hand)
"""
if max_left_hand is None:
max_left_hand = max_per_hand
midi_out = copy.deepcopy(midi_data)
trimmed = 0
for instrument in midi_out.instruments:
notes = sorted(instrument.notes, key=lambda n: n.start)
if not notes:
continue
for i, note in enumerate(notes):
is_right = note.pitch >= hand_split
limit = max_per_hand if is_right else max_left_hand
# Find all notes in the same hand currently sounding
sounding = []
for j in range(i):
if notes[j].end > note.start:
same_hand = (notes[j].pitch >= hand_split) == is_right
if same_hand:
sounding.append(j)
if len(sounding) + 1 > limit:
excess = len(sounding) + 1 - limit
all_indices = sounding + [i]
# Both hands: protect highest pitch (melody voice)
protected = max(all_indices, key=lambda j: notes[j].pitch)
# Among the sustained (not the new note), trim lowest velocity
# but never trim the protected note
trimmable = [j for j in sounding if j != protected]
scored = [(notes[j].velocity, j) for j in trimmable]
scored.sort() # lowest velocity trimmed first
for _, j in scored[:excess]:
notes[j].end = note.start
trimmed += 1
instrument.notes = [n for n in notes if n.end - n.start > 0.01]
return midi_out, trimmed
def remove_hand_outliers(midi_data, hand_split=60, gap_threshold=7):
"""Remove notes that are pitch outliers within their hand group.
For each chord (notes within 30ms), splits into left/right hand and
checks for notes isolated from the cluster at the low end — e.g. a
left-hand note at MIDI 33 when the rest of the LH chord is at 45-52,
or a right-hand note at MIDI 62 when the rest is at 72-79.
Both hands protect the melody (highest note) and flag the lowest note
as an outlier if it's too far from the cluster. These low outliers are
almost always sub-harmonic ghosts from the transcriber.
Args:
hand_split: MIDI pitch dividing left/right hand (default 60 = C4)
gap_threshold: Semitones — if a note is this far from its nearest
neighbor in the same hand, it's flagged as an outlier (default 7)
"""
midi_out = copy.deepcopy(midi_data)
removed = 0
for instrument in midi_out.instruments:
notes = sorted(instrument.notes, key=lambda n: n.start)
if not notes:
continue
# Group into chords (notes within 30ms)
chords = []
current_chord = [0]
for i in range(1, len(notes)):
if notes[i].start - notes[current_chord[0]].start < 0.03:
current_chord.append(i)
else:
chords.append(current_chord)
current_chord = [i]
chords.append(current_chord)
to_remove = set()
for chord_indices in chords:
left = [idx for idx in chord_indices if notes[idx].pitch < hand_split]
right = [idx for idx in chord_indices if notes[idx].pitch >= hand_split]
for hand_indices in [right, left]:
if len(hand_indices) < 3:
# Need at least 3 notes to identify an outlier vs cluster
continue
pitches = sorted([(notes[idx].pitch, idx) for idx in hand_indices])
# Both hands: melody (highest) is protected.
# Check if the lowest note is far from the cluster.
lowest_pitch, lowest_idx = pitches[0]
second_pitch = pitches[1][0]
gap = second_pitch - lowest_pitch
if gap >= gap_threshold:
to_remove.add(lowest_idx)
instrument.notes = [n for k, n in enumerate(notes) if k not in to_remove]
removed += len(to_remove)
return midi_out, removed
def enforce_hand_span(midi_data, max_span=12, hand_split=60):
"""Enforce that no hand plays notes wider than max_span semitones.
Both hands anchor on the MELODY (highest note) and build downward.
This matches real piano technique: the top voice carries the melody
and harmonics are voiced below within reach.
Checks both:
1. Chord groups (notes starting within 30ms)
2. Concurrent sounding notes (sustained notes overlapping new ones)
For LH: protects highest note (melody line), removes lowest that
exceed the span — the melody voice is the most important.
For RH: protects highest note (melody), removes lowest.
Args:
max_span: Maximum interval in semitones (default 12 = octave)
hand_split: MIDI pitch dividing left/right hand (default 60 = C4)
"""
midi_out = copy.deepcopy(midi_data)
removed = 0
for instrument in midi_out.instruments:
notes = sorted(instrument.notes, key=lambda n: n.start)
if not notes:
continue
# ── Pass 1: Chord groups (simultaneous onsets within 30ms) ──
chords = []
current_chord = [0]
for i in range(1, len(notes)):
if notes[i].start - notes[current_chord[0]].start < 0.03:
current_chord.append(i)
else:
chords.append(current_chord)
current_chord = [i]
chords.append(current_chord)
to_remove = set()
for chord_indices in chords:
left = [idx for idx in chord_indices if notes[idx].pitch < hand_split]
right = [idx for idx in chord_indices if notes[idx].pitch >= hand_split]
for hand_indices in [right, left]:
if len(hand_indices) < 2:
continue
pitches = sorted(hand_indices, key=lambda idx: notes[idx].pitch)
span = notes[pitches[-1]].pitch - notes[pitches[0]].pitch
if span <= max_span:
continue
# Both hands: protect highest (melody), remove lowest
anchor_pitch = notes[pitches[-1]].pitch
for idx in pitches[:-1]:
if anchor_pitch - notes[idx].pitch > max_span:
to_remove.add(idx)
# ── Pass 2: Concurrent sounding notes (sustained overlap) ──
for i, note in enumerate(notes):
if i in to_remove:
continue
is_right = note.pitch >= hand_split
# Find all same-hand notes currently sounding
concurrent = [i]
for j in range(i):
if j in to_remove:
continue
if notes[j].end > note.start + 0.01:
if (notes[j].pitch >= hand_split) == is_right:
concurrent.append(j)
if len(concurrent) < 2:
continue
pitches_conc = sorted(concurrent, key=lambda idx: notes[idx].pitch)
span = notes[pitches_conc[-1]].pitch - notes[pitches_conc[0]].pitch
if span <= max_span:
continue
# Protect highest (melody), trim lowest sustained notes
anchor_pitch = notes[pitches_conc[-1]].pitch
for idx in pitches_conc[:-1]:
if anchor_pitch - notes[idx].pitch > max_span:
# Don't remove entirely — just end the sustained note
notes[idx].end = note.start
if notes[idx].end - notes[idx].start < 0.05:
to_remove.add(idx)
removed += 1
instrument.notes = [n for k, n in enumerate(notes) if k not in to_remove]
removed += len(to_remove)
return midi_out, removed
def merge_repeated_notes(midi_data, y, sr, hop_length=512, min_gap=0.15):
"""Merge consecutive same-pitch notes that lack a real re-attack.
Basic-pitch often fragments a single sustained note into multiple short
re-strikes. This step checks whether a repeated note has genuine onset
energy at the re-attack point. If not, the notes are merged into one
sustained note.
Args:
min_gap: If the gap between notes is larger than this (seconds),
always keep separate — the silence itself is musical. Default 150ms.
"""
midi_out = copy.deepcopy(midi_data)
merged_count = 0
# Compute onset strength envelope for verification
onset_env = librosa.onset.onset_strength(y=y, sr=sr, hop_length=hop_length)
for instrument in midi_out.instruments:
# Sort by pitch then start time to find consecutive same-pitch notes
notes = sorted(instrument.notes, key=lambda n: (n.pitch, n.start))
to_remove = set()
i = 0
while i < len(notes) - 1:
if i in to_remove:
i += 1
continue
note = notes[i]
j = i + 1
# Walk forward through consecutive same-pitch notes
while j < len(notes) and notes[j].pitch == note.pitch:
if j in to_remove:
j += 1
continue
next_note = notes[j]
gap = next_note.start - note.end
# If there's a real gap (silence), keep them separate
if gap > min_gap:
break
# If the next note starts before or just after this one ends,
# check for onset energy at the re-attack point
reattack_time = next_note.start
reattack_frame = int(reattack_time * sr / hop_length)
has_onset = False
if 0 <= reattack_frame < len(onset_env):
# Check onset strength in a small window around the re-attack
lo = max(0, reattack_frame - 1)
hi = min(len(onset_env), reattack_frame + 2)
local_strength = float(np.max(onset_env[lo:hi]))
# Compare to the median onset strength — if re-attack is
# weaker than median, it's not a real new attack
median_strength = float(np.median(onset_env[onset_env > 0])) if np.any(onset_env > 0) else 0
has_onset = local_strength > median_strength * 2.0
if not has_onset:
# Merge: extend current note to cover the next one
note.end = max(note.end, next_note.end)
to_remove.add(j)
merged_count += 1
j += 1
else:
# Real re-attack — stop merging
break
i = j if j > i + 1 else i + 1
instrument.notes = [n for k, n in enumerate(notes) if k not in to_remove]
return midi_out, merged_count
def consolidate_rhythm(midi_data, y, sr, hop_length=512, max_snap=0.06):
"""Consolidate note onsets onto a dominant rhythmic pattern.
After onset correction, notes can scatter across many different
micro-timings, losing the clean rhythmic feel. This step:
1. Detects tempo and beat positions
2. Builds a histogram of note positions within each beat (16 bins
per beat = 16th-note resolution)
3. Identifies dominant subdivisions (top positions by note count,
capped at 8 max)
4. Re-snaps all onsets to the nearest dominant subdivision
Onsets already on a dominant position are untouched. Stray onsets
are snapped only if within max_snap seconds of a dominant position.
Args:
max_snap: Maximum distance to snap (default 60ms). Notes further
from any dominant position are left alone.
"""
midi_out = copy.deepcopy(midi_data)
# Detect tempo and beats
tempo, beat_frames = librosa.beat.beat_track(y=y, sr=sr, hop_length=hop_length)
if hasattr(tempo, '__len__'):
tempo = float(tempo[0])
# Fix tempo doubling
if tempo > 140:
half_tempo = tempo / 2
if 50 <= half_tempo <= 120:
tempo = half_tempo
beat_frames = beat_frames[::2]
beat_times = librosa.frames_to_time(beat_frames, sr=sr, hop_length=hop_length)
if len(beat_times) < 4:
return midi_out, 0, 0
# Collect all note onsets
all_notes = []
for inst_idx, inst in enumerate(midi_out.instruments):
for note in inst.notes:
all_notes.append(note)
if not all_notes:
return midi_out, 0, 0
# ── Step 1: Build histogram of where notes fall within each beat ──
# Use 16 bins per beat (16th-note resolution)
n_bins = 16
histogram = np.zeros(n_bins)
for note in all_notes:
# Find which beat this note belongs to
beat_idx = np.searchsorted(beat_times, note.start, side='right') - 1
if beat_idx < 0 or beat_idx >= len(beat_times) - 1:
continue
beat_start = beat_times[beat_idx]
beat_dur = beat_times[beat_idx + 1] - beat_start
if beat_dur <= 0:
continue
# Position within beat as fraction [0, 1)
frac = (note.start - beat_start) / beat_dur
frac = max(0.0, min(frac, 0.9999))
bin_idx = int(frac * n_bins)
histogram[bin_idx] += 1
total_notes_in_beats = histogram.sum()
if total_notes_in_beats == 0:
return midi_out, 0, 0
# ── Step 2: Identify dominant subdivisions ──
# Pick the top bins by note count. Always include downbeat (0) and
# half-beat (8). Cap at 8 dominant positions max to force a clean grid.
dominant_bins = {0}
if histogram[8] > 0:
dominant_bins.add(8)
# Sort bins by count (descending), add until we have up to 4
# Fewer dominant positions = tighter grid = cleaner rhythm
ranked = sorted(range(n_bins), key=lambda i: histogram[i], reverse=True)
min_count = max(total_notes_in_beats * 0.05, 4) # must have at least 5% or 4 notes
for b in ranked:
if len(dominant_bins) >= 4:
break
if histogram[b] >= min_count:
dominant_bins.add(b)
dominant_fracs = sorted([b / n_bins for b in dominant_bins])
print(f" Dominant subdivisions: {len(dominant_fracs)}/{n_bins} "
f"(bins: {sorted(dominant_bins)})")
# ── Step 3: Build full grid of dominant positions ──
dominant_grid = []
for i in range(len(beat_times) - 1):
beat_start = beat_times[i]
beat_dur = beat_times[i + 1] - beat_start
for frac in dominant_fracs:
dominant_grid.append(beat_start + frac * beat_dur)
# Extend past the last beat
if len(beat_times) >= 2:
last_dur = beat_times[-1] - beat_times[-2]
for frac in dominant_fracs:
dominant_grid.append(beat_times[-1] + frac * last_dur)
dominant_grid = np.array(dominant_grid)
# ── Step 4: Build 8th-note fallback grid ──
# For notes that are too far from any dominant position, snap to the
# nearest 8th note instead of leaving them unquantized.
beat_dur = 60.0 / tempo if tempo > 30 else 0.5
eighth = beat_dur / 2.0
fallback_grid = []
if len(beat_times) >= 2:
fb_start = max(0, beat_times[0] - beat_dur * 2)
fb_t = fb_start
while fb_t <= beat_times[-1] + beat_dur * 2:
fallback_grid.append(fb_t)
fb_t += eighth
fallback_grid = np.array(fallback_grid) if fallback_grid else np.array([0])
# ── Step 5: Snap stray onsets to dominant grid (or fallback) ──
snapped = 0
for inst in midi_out.instruments:
for note in inst.notes:
diffs = np.abs(dominant_grid - note.start)
nearest_idx = np.argmin(diffs)
dist = diffs[nearest_idx]
if dist < 0.003:
# Already on a dominant position (within 3ms)
continue
if dist <= max_snap:
duration = note.end - note.start
note.start = dominant_grid[nearest_idx]
note.end = note.start + duration
snapped += 1
else:
# Fallback: snap to nearest 8th note
fb_diffs = np.abs(fallback_grid - note.start)
fb_idx = np.argmin(fb_diffs)
if fb_diffs[fb_idx] <= max_snap * 1.2:
duration = note.end - note.start
note.start = fallback_grid[fb_idx]
note.end = note.start + duration
snapped += 1
return midi_out, snapped, len(dominant_fracs)
def detect_sustain_regions(y, sr, hop_length=512):
"""Detect regions where the sustain pedal is likely engaged.
Analyzes spectral flux and broadband energy decay. When the sustain pedal
is held, notes ring longer and the spectral energy decays slowly instead
of dropping abruptly at note release. Detects this by looking for:
1. Low spectral flux (sustained timbre, no new attacks)
2. Slow energy decay (notes ringing through pedal)
Returns a boolean array (per frame) indicating sustained regions.
"""
# Compute spectral flux (rate of spectral change)
S = np.abs(librosa.stft(y, hop_length=hop_length))
flux = np.sqrt(np.mean(np.diff(S, axis=1) ** 2, axis=0))
flux = np.concatenate([[0], flux]) # pad to match frame count
# Compute RMS energy
rms = librosa.feature.rms(y=y, hop_length=hop_length)[0]
# Normalize both
flux_norm = flux / (np.percentile(flux, 95) + 1e-8)
rms_norm = rms / (np.max(rms) + 1e-8)
n_frames = min(len(flux_norm), len(rms_norm))
flux_norm = flux_norm[:n_frames]
rms_norm = rms_norm[:n_frames]
# Sustain pedal indicators:
# - Low spectral flux (< 30th percentile) = sustained sound, not new attacks
# - Moderate+ energy (> 10% of peak) = notes are still ringing
flux_thresh = np.percentile(flux_norm, 30)
sustain_mask = (flux_norm < flux_thresh) & (rms_norm > 0.10)
# Smooth: close 200ms gaps, remove blips shorter than 300ms
close_frames = max(1, int(0.2 * sr / hop_length))
min_region = max(1, int(0.3 * sr / hop_length))
# Morphological closing
for i in range(1, n_frames - 1):
if not sustain_mask[i]:
before = any(sustain_mask[max(0, i - close_frames):i])
after = any(sustain_mask[i + 1:min(n_frames, i + close_frames + 1)])
if before and after:
sustain_mask[i] = True
# Remove short blips
in_region = False
start = 0
for i in range(n_frames):
if sustain_mask[i] and not in_region:
start = i
in_region = True
elif not sustain_mask[i] and in_region:
if i - start < min_region:
sustain_mask[start:i] = False
in_region = False
return sustain_mask
def extend_note_durations(midi_data, y, sr, hop_length=512, max_per_hand=4, hand_split=60):
"""Extend MIDI note durations to match audio CQT energy decay.
Basic-pitch systematically underestimates note durations. This uses
the CQT spectrogram to find where the audio energy actually decays
and extends each note to match, dramatically improving spectral recall.
Concurrent-aware: won't extend a note past the point where doing so
would exceed max_per_hand concurrent notes in the same hand. This
prevents the downstream concurrent limiter from having to trim hundreds
of over-extended notes.
"""
midi_out = copy.deepcopy(midi_data)
N_BINS = 88 * 3
FMIN = librosa.note_to_hz('A0')
C = np.abs(librosa.cqt(
y, sr=sr, hop_length=hop_length,
fmin=FMIN, n_bins=N_BINS, bins_per_octave=36,
))
C_db = librosa.amplitude_to_db(C, ref=np.max(C))
C_norm = np.maximum(C_db, -80.0)
C_norm = (C_norm + 80.0) / 80.0
n_frames = C.shape[1]
# Detect sustain pedal regions for longer extension allowance
sustain_mask = detect_sustain_regions(y, sr, hop_length)
# Pad/trim to match CQT frame count
if len(sustain_mask) < n_frames:
sustain_mask = np.concatenate([sustain_mask, np.zeros(n_frames - len(sustain_mask), dtype=bool)])
else:
sustain_mask = sustain_mask[:n_frames]
# Pre-compute per-frame concurrent counts per hand (fast O(1) lookup)
right_count = np.zeros(n_frames, dtype=int)
left_count = np.zeros(n_frames, dtype=int)
for inst in midi_out.instruments:
for n in inst.notes:
sf = max(0, int(n.start * sr / hop_length))
ef = min(n_frames, int(n.end * sr / hop_length))
if n.pitch >= hand_split:
right_count[sf:ef] += 1
else:
left_count[sf:ef] += 1
extended = 0
sustain_extended = 0
for inst in midi_out.instruments:
# Sort notes by start time for overlap checking
notes_sorted = sorted(inst.notes, key=lambda n: (n.pitch, n.start))
for idx, note in enumerate(notes_sorted):
fund_bin = (note.pitch - 21) * 3 + 1
if fund_bin < 0 or fund_bin >= N_BINS:
continue
end_frame = min(n_frames, int(note.end * sr / hop_length))
# In sustain regions, allow longer extension (4s) and lower threshold
in_sustain = end_frame < n_frames and sustain_mask[min(end_frame, n_frames - 1)]
max_ext_seconds = 4.0 if in_sustain else 2.0
energy_thresh = 0.15 if in_sustain else 0.20
max_extend = min(n_frames, end_frame + int(max_ext_seconds * sr / hop_length))
# Don't extend into the next note at the same pitch
next_start_frame = max_extend
for other in notes_sorted[idx + 1:]:
if other.pitch == note.pitch:
next_start_frame = min(next_start_frame, int(other.start * sr / hop_length) - 1)
break
is_right = note.pitch >= hand_split
hand_count = right_count if is_right else left_count
actual_end = end_frame
for f in range(end_frame, min(max_extend, next_start_frame)):
lo = max(0, fund_bin - 1)
hi = min(N_BINS, fund_bin + 2)
if np.mean(C_norm[lo:hi, f]) > energy_thresh:
# Check concurrent: this note isn't counted in hand_count
# beyond end_frame, so hand_count[f] >= max_per_hand means
# extending here would create max_per_hand + 1 concurrent
if hand_count[f] >= max_per_hand:
break
actual_end = f
else:
break
new_end = actual_end * hop_length / sr
if new_end > note.end + 0.05:
# Update the concurrent count array for the extended region
old_end_frame = end_frame
new_end_frame = min(n_frames, int(new_end * sr / hop_length))
if new_end_frame > old_end_frame:
hand_count[old_end_frame:new_end_frame] += 1
note.end = new_end
extended += 1
if in_sustain:
sustain_extended += 1
return midi_out, extended
def align_chords(midi_data, threshold=0.02):
"""Snap notes within a chord to the exact same onset time.
basic-pitch's ~12ms frame resolution can make notes in the same chord
start at slightly different times, causing a 'flammy' sound.
"""
midi_out = copy.deepcopy(midi_data)
aligned = 0
for instrument in midi_out.instruments:
notes = sorted(instrument.notes, key=lambda n: n.start)
i = 0
while i < len(notes):
group = [notes[i]]
j = i + 1
while j < len(notes) and notes[j].start - notes[i].start < threshold:
group.append(notes[j])
j += 1
if len(group) > 1:
median_start = float(np.median([n.start for n in group]))
for note in group:
if note.start != median_start:
duration = note.end - note.start
note.start = median_start
note.end = median_start + duration
aligned += 1
i = j
return midi_out, aligned
def quantize_to_beat_grid(midi_data, y, sr, hop_length=512, strength=0.5):
"""Quantize note onsets to a detected beat grid.
Uses librosa beat tracking to find the tempo and beat positions,
builds a 16th-note grid, and snaps onsets toward the nearest grid
position. Preserves note durations.
"""
midi_out = copy.deepcopy(midi_data)
tempo, beat_frames = librosa.beat.beat_track(y=y, sr=sr, hop_length=hop_length)
if hasattr(tempo, '__len__'):
tempo = float(tempo[0])
# Fix tempo doubling: librosa often detects double the true tempo for
# slow/moderate songs (e.g., 86 BPM → 172). If tempo > 140 and halving
# gives a reasonable tempo (50-120), use the half tempo and keep only
# every other beat.
if tempo > 140:
half_tempo = tempo / 2
if 50 <= half_tempo <= 120:
tempo = half_tempo
beat_frames = beat_frames[::2] # keep every other beat
beat_times = librosa.frames_to_time(beat_frames, sr=sr, hop_length=hop_length)
if len(beat_times) < 2:
print(" Could not detect beats, skipping quantization")
return midi_out, 0, tempo
# Build a 16th-note grid from the beat times
grid = []
for i in range(len(beat_times) - 1):
beat_dur = beat_times[i + 1] - beat_times[i]
sixteenth = beat_dur / 4
for sub in range(4):
grid.append(beat_times[i] + sub * sixteenth)
if len(beat_times) >= 2:
last_beat_dur = beat_times[-1] - beat_times[-2]
sixteenth = last_beat_dur / 4
for sub in range(4):
grid.append(beat_times[-1] + sub * sixteenth)
grid = np.array(grid)
quantized = 0
for instrument in midi_out.instruments:
for note in instrument.notes:
diffs = np.abs(grid - note.start)
nearest_idx = np.argmin(diffs)
nearest_grid = grid[nearest_idx]
deviation = nearest_grid - note.start
if abs(deviation) < 0.06: # Only quantize if < 60ms off grid
duration = note.end - note.start
note.start = note.start + deviation * strength
note.end = note.start + duration
if abs(deviation) > 0.005:
quantized += 1
return midi_out, quantized, tempo
def correct_onsets(midi_data, ref_onsets, min_off=0.02, max_off=0.15):
"""Correct chord onsets that are clearly misaligned with audio onsets.
Groups notes into chords, then for each chord checks if there's a closer
audio onset. Only corrects if min_off-max_off off and no adjacent chord
is a better match.
"""
midi_out = copy.deepcopy(midi_data)
all_notes = sorted(
[(n, inst_idx) for inst_idx, inst in enumerate(midi_out.instruments)
for n in inst.notes],
key=lambda x: x[0].start
)
chord_groups = []
if all_notes:
current_group = [all_notes[0]]
for item in all_notes[1:]:
if item[0].start - current_group[0][0].start < 0.03:
current_group.append(item)
else:
chord_groups.append(current_group)
current_group = [item]
chord_groups.append(current_group)
chord_onsets = np.array([g[0][0].start for g in chord_groups])
corrections = 0
total_shift = 0.0
for group_idx, group in enumerate(chord_groups):
chord_onset = chord_onsets[group_idx]
diffs = ref_onsets - chord_onset
abs_diffs = np.abs(diffs)
nearest_idx = np.argmin(abs_diffs)
nearest_diff = diffs[nearest_idx]
abs_diff = abs_diffs[nearest_idx]
if min_off < abs_diff < max_off:
# Verify no adjacent chord is a better match
if group_idx > 0:
prev_onset = chord_onsets[group_idx - 1]
if abs(ref_onsets[nearest_idx] - prev_onset) < abs_diff:
continue
if group_idx < len(chord_onsets) - 1:
next_onset = chord_onsets[group_idx + 1]
if abs(ref_onsets[nearest_idx] - next_onset) < abs_diff:
continue
for note, inst_idx in group:
duration = note.end - note.start
note.start = max(0, note.start + nearest_diff)
note.end = note.start + duration
corrections += 1
total_shift += abs(nearest_diff)
initial_f1 = onset_f1(ref_onsets, chord_onsets)
corrected_onsets = np.array([g[0][0].start for g in chord_groups])
final_f1 = onset_f1(ref_onsets, corrected_onsets)
return midi_out, corrections, total_shift, len(chord_groups), initial_f1, final_f1
def apply_global_offset(midi_data, ref_onsets):
"""Measure and correct systematic timing offset against audio onsets.
Computes the median difference between MIDI and audio onsets, then
shifts all notes to center the distribution around zero.
"""
midi_out = copy.deepcopy(midi_data)
all_onsets = sorted(set(n.start for inst in midi_out.instruments for n in inst.notes))
diffs = []
for mo in all_onsets:
ad = np.abs(ref_onsets - mo)
if np.min(ad) < 0.10:
closest = ref_onsets[np.argmin(ad)]
diffs.append(closest - mo) # positive = MIDI is early, negative = late
if not diffs:
return midi_out, 0.0
median_offset = float(np.median(diffs))
# Only apply if the offset is meaningful (> 5ms)
if abs(median_offset) < 0.005:
return midi_out, 0.0
for instrument in midi_out.instruments:
for note in instrument.notes:
duration = note.end - note.start
note.start = max(0, note.start + median_offset)
note.end = note.start + duration
return midi_out, median_offset
def fix_note_overlap(midi_data, hand_split=60, min_duration=0.10):
"""Trim overlapping notes in the right hand so each note releases cleanly.
Also enforces a minimum note duration across ALL notes.
"""
midi_out = copy.deepcopy(midi_data)
trimmed = 0
for instrument in midi_out.instruments:
rh_notes = [n for n in instrument.notes if n.pitch >= hand_split]
rh_notes.sort(key=lambda n: (n.start, n.pitch))
for i, note in enumerate(rh_notes):
for j in range(i + 1, min(i + 8, len(rh_notes))):
next_note = rh_notes[j]
if next_note.start <= note.start:
continue
overlap = note.end - next_note.start
if overlap > 0.05: # Only trim significant overlaps (>50ms)
original_dur = note.end - note.start
new_end = next_note.start - 0.01
# Never shorten more than 30% of original duration
min_allowed = note.start + original_dur * 0.7
if new_end < min_allowed:
new_end = min_allowed
note.end = new_end
if note.end - note.start < min_duration:
note.end = note.start + min_duration
trimmed += 1
break
# Enforce minimum duration on ALL notes (catches any collapsed durations)
enforced = 0
for instrument in midi_out.instruments:
for note in instrument.notes:
if note.end - note.start < min_duration:
note.end = note.start + min_duration
enforced += 1
return midi_out, trimmed, enforced
def recover_missing_notes(midi_data, y, sr, hop_length=512, snap_onsets=None):
"""Recover strong notes the transcriber missed using CQT analysis.
Scans the audio CQT for pitch energy that isn't represented in the MIDI.
When a pitch has strong, sustained energy but no corresponding MIDI note,
synthesize one. Targets upper register (>= C4) where basic-pitch can
under-detect, especially when harmonics cause false removal.
If snap_onsets is provided, recovered notes are snapped to the nearest
existing onset for rhythmic alignment.
Should be run AFTER all removal filters so the coverage map reflects
what actually survived.
"""
midi_out = copy.deepcopy(midi_data)
N_BINS = 88 * 3
FMIN = librosa.note_to_hz('A0')
C = np.abs(librosa.cqt(
y, sr=sr, hop_length=hop_length,
fmin=FMIN, n_bins=N_BINS, bins_per_octave=36,
))
C_db = librosa.amplitude_to_db(C, ref=np.max(C))
times = librosa.frames_to_time(np.arange(C.shape[1]), sr=sr, hop_length=hop_length)
# Build a set of existing note coverage: (pitch, frame) pairs
existing = set()
for inst in midi_out.instruments:
for note in inst.notes:
start_frame = max(0, int(note.start * sr / hop_length))
end_frame = min(C.shape[1], int(note.end * sr / hop_length))
for f in range(start_frame, end_frame):
existing.add((note.pitch, f))
# Scan C4 (60) to B6 (95) for uncovered energy
recovered = 0
min_energy = -10.0 # dB threshold — only recover notes with strong CQT energy
min_duration_s = 0.05 # ~50ms minimum
gap_tolerance = 3 # allow 3-frame dips without breaking a note
for midi_pitch in range(60, 96):
fund_bin = (midi_pitch - 21) * 3 + 1
if fund_bin < 0 or fund_bin >= N_BINS:
continue
# Harmonic check: skip if an octave-below note is much louder
# (this note is likely a harmonic, not a real played note)
lower_pitch = midi_pitch - 12
if lower_pitch >= 21:
lower_bin = (lower_pitch - 21) * 3 + 1
if 0 <= lower_bin < N_BINS:
lower_lo = max(0, lower_bin - 1)
lower_hi = min(N_BINS, lower_bin + 2)
upper_energy = float(np.max(C_db[max(0, fund_bin - 1):min(N_BINS, fund_bin + 2), :]))
lower_energy = float(np.max(C_db[lower_lo:lower_hi, :]))
if lower_energy - upper_energy > 12:
# Octave below is 12+ dB louder — likely a harmonic
continue
lo = max(0, fund_bin - 1)
hi = min(N_BINS, fund_bin + 2)
# Get energy and coverage for this pitch
pitch_energy = np.max(C_db[lo:hi, :], axis=0)
# Find uncovered regions with strong energy
strong_uncovered = np.array([
pitch_energy[f] >= min_energy and (midi_pitch, f) not in existing
for f in range(len(pitch_energy))
])
# Close small gaps (morphological closing)
for f in range(1, len(strong_uncovered) - 1):
if not strong_uncovered[f] and pitch_energy[f] >= min_energy - 5:
before = any(strong_uncovered[max(0, f - gap_tolerance):f])
after = any(strong_uncovered[f + 1:min(len(strong_uncovered), f + gap_tolerance + 1)])
if before and after:
strong_uncovered[f] = True
# Extract contiguous regions
regions = []
in_region = False
start_f = 0
for f in range(len(strong_uncovered)):
if strong_uncovered[f] and not in_region:
start_f = f
in_region = True
elif not strong_uncovered[f] and in_region:
regions.append((start_f, f))
in_region = False
if in_region:
regions.append((start_f, len(strong_uncovered)))
for start_f, end_f in regions:
t_start = times[start_f]
t_end = times[min(end_f, len(times) - 1)]
if t_end - t_start < min_duration_s:
continue
avg_energy = float(np.mean(pitch_energy[start_f:end_f]))
velocity = min(75, max(35, int(55 + avg_energy * 1.5)))
# Snap to nearest existing onset for rhythmic alignment
note_start = t_start
note_end = t_end
if snap_onsets is not None and len(snap_onsets) > 0:
snap_arr = np.array(snap_onsets)
diffs = np.abs(snap_arr - t_start)
nearest_idx = np.argmin(diffs)
if diffs[nearest_idx] < 0.06:
dur = t_end - t_start
note_start = snap_arr[nearest_idx]
note_end = note_start + dur
new_note = pretty_midi.Note(
velocity=velocity,
pitch=midi_pitch,
start=note_start,
end=note_end,
)
midi_out.instruments[0].notes.append(new_note)
recovered += 1
return midi_out, recovered
def estimate_complexity(midi_data, audio_duration):
"""Estimate piece complexity to adjust filter aggressiveness.
Returns a dict with:
- note_density: notes per second
- avg_polyphony: average concurrent notes at any onset
- complexity: 'simple' (<4 n/s), 'moderate' (4-8), 'complex' (>8)
Complex pieces need less aggressive ghost removal and wider tolerance
for concurrent notes, since dense textures are intentional.
"""
all_notes = sorted(
[n for inst in midi_data.instruments for n in inst.notes],
key=lambda n: n.start
)
if not all_notes or audio_duration < 1:
return {'note_density': 0, 'avg_polyphony': 1, 'complexity': 'simple'}
note_density = len(all_notes) / audio_duration
# Compute average polyphony at each onset
onsets = sorted(set(round(n.start, 3) for n in all_notes))
polyphonies = []
for onset in onsets:
count = sum(1 for n in all_notes if abs(n.start - onset) < 0.03)
polyphonies.append(count)
avg_polyphony = np.mean(polyphonies) if polyphonies else 1
if note_density > 8 or avg_polyphony > 3.5:
complexity = 'complex'
elif note_density > 4 or avg_polyphony > 2.5:
complexity = 'moderate'
else:
complexity = 'simple'
return {
'note_density': note_density,
'avg_polyphony': avg_polyphony,
'complexity': complexity,
}
def optimize(original_audio_path, midi_path, output_path=None):
"""Full optimization pipeline."""
if output_path is None:
output_path = midi_path
sr = 22050
hop_length = 512
# Load audio and detect onsets
print(f"Analyzing audio: {original_audio_path}")
y, _ = librosa.load(original_audio_path, sr=sr, mono=True)
audio_duration = len(y) / sr
onset_env = librosa.onset.onset_strength(y=y, sr=sr, hop_length=hop_length)
# Use backtrack=False: basic-pitch onsets align with energy peaks, not
# the earlier rise points that backtrack finds (~50ms earlier).
# Use delta=0.04 for higher sensitivity — detects ~15% more onsets,
# reducing unmatched MIDI onsets from 116 to 80.
ref_onset_frames = librosa.onset.onset_detect(
onset_envelope=onset_env, sr=sr, hop_length=hop_length,
backtrack=False, delta=0.04
)
ref_onsets = librosa.frames_to_time(ref_onset_frames, sr=sr, hop_length=hop_length)
print(f" {audio_duration:.1f}s, {len(ref_onsets)} audio onsets")
# Load MIDI
midi_data = pretty_midi.PrettyMIDI(str(midi_path))
total_notes = sum(len(inst.notes) for inst in midi_data.instruments)
print(f" {total_notes} MIDI notes")
# Estimate complexity to adjust filter thresholds
complexity_info = estimate_complexity(midi_data, audio_duration)
complexity = complexity_info['complexity']
print(f" Complexity: {complexity} (density={complexity_info['note_density']:.1f} n/s, "
f"polyphony={complexity_info['avg_polyphony']:.1f})")
# Step 0: Remove notes in leading silence (mic rumble artifacts)
print("\nStep 0: Removing notes in leading silence...")
midi_data, silence_removed, music_start = remove_leading_silence_notes(midi_data, y, sr)
if silence_removed:
print(f" Music starts at {music_start:.2f}s, removed {silence_removed} noise notes")
else:
print(f" No leading silence detected")
# Step 0b: Remove notes in trailing silence
print("\nStep 0b: Removing notes in trailing silence...")
midi_data, trail_removed, music_end = remove_trailing_silence_notes(midi_data, y, sr)
if trail_removed:
print(f" Music ends at {music_end:.2f}s, removed {trail_removed} trailing notes")
else:
print(f" No trailing silence notes detected")
# Step 0c: Remove low-energy hallucinations
print("\nStep 0c: Removing low-energy hallucinations...")
midi_data, energy_removed = remove_low_energy_notes(midi_data, y, sr, hop_length)
print(f" Removed {energy_removed} notes with no audio onset energy")
# Step 0d: Remove harmonic ghost notes (CQT-aware)
print("\nStep 0d: Removing harmonic ghost notes...")
midi_data, ghosts_removed = remove_harmonic_ghosts(midi_data, y, sr, hop_length)
print(f" Removed {ghosts_removed} octave-harmonic ghosts")
# Step 1: Remove phantom high notes (conservative)
print("\nStep 1: Removing phantom high notes...")
midi_data, phantoms_removed = remove_phantom_notes(midi_data)
print(f" Removed {phantoms_removed} phantom notes")
# Step 1b: Hard pitch ceiling at C7 (MIDI 96) — extreme highs only
print("\nStep 1b: Applying pitch ceiling (C7 / MIDI 96)...")
midi_data, ceiling_removed = apply_pitch_ceiling(midi_data, max_pitch=96)
print(f" Removed {ceiling_removed} notes above C7")
# Step 2: Align chord notes to single onset
print("\nStep 2: Aligning chord notes...")
midi_data, chords_aligned = align_chords(midi_data)
print(f" Aligned {chords_aligned} notes within chords")
# Step 3: Full beat-grid quantization
print("\nStep 3: Quantizing to beat grid...")
midi_data, notes_quantized, detected_tempo = quantize_to_beat_grid(
midi_data, y, sr, hop_length, strength=1.0
)
print(f" Detected tempo: {detected_tempo:.0f} BPM")
print(f" Quantized {notes_quantized} notes (full snap)")
# Step 4: Targeted onset correction against audio
print("\nStep 4: Correcting onsets against audio...")
midi_data, corrections, total_shift, n_chords, pre_f1, post_f1 = \
correct_onsets(midi_data, ref_onsets)
avg_shift = (total_shift / corrections * 1000) if corrections > 0 else 0
print(f" Corrected {corrections}/{n_chords} (avg {avg_shift:.0f}ms)")
print(f" Onset F1: {pre_f1:.4f} -> {post_f1:.4f}")
# Step 5: Tight second correction pass (10-60ms window)
print("\nStep 5: Fine-tuning onsets (tight pass)...")
midi_data, corrections2, total_shift2, n_chords2, _, post_f1_2 = \
correct_onsets(midi_data, ref_onsets, min_off=0.01, max_off=0.06)
avg_shift2 = (total_shift2 / corrections2 * 1000) if corrections2 > 0 else 0
print(f" Fine-tuned {corrections2}/{n_chords2} (avg {avg_shift2:.0f}ms)")
print(f" Onset F1: {post_f1:.4f} -> {post_f1_2:.4f}")
# Step 6: Micro-correction pass (5-25ms window)
print("\nStep 6: Micro-correcting onsets...")
midi_data, corrections3, total_shift3, n_chords3, _, post_f1_3 = \
correct_onsets(midi_data, ref_onsets, min_off=0.005, max_off=0.025)
avg_shift3 = (total_shift3 / corrections3 * 1000) if corrections3 > 0 else 0
print(f" Micro-corrected {corrections3}/{n_chords3} (avg {avg_shift3:.0f}ms)")
print(f" Onset F1: {post_f1_2:.4f} -> {post_f1_3:.4f}")
# Step 6b: Remove spurious false-positive onsets
print("\nStep 6b: Removing spurious onsets (false positive cleanup)...")
midi_data, spurious_notes, spurious_onsets = remove_spurious_onsets(
midi_data, y, sr, ref_onsets, hop_length, complexity=complexity
)
print(f" Removed {spurious_notes} notes across {spurious_onsets} spurious onsets")
# Step 6c: Wide onset recovery pass (50-120ms window) to rescue false negatives
print("\nStep 6c: Wide onset recovery (rescuing false negatives)...")
midi_data, corrections_wide, total_shift_wide, n_chords_wide, _, post_f1_wide = \
correct_onsets(midi_data, ref_onsets, min_off=0.04, max_off=0.12)
avg_shift_wide = (total_shift_wide / corrections_wide * 1000) if corrections_wide > 0 else 0
print(f" Recovered {corrections_wide}/{n_chords_wide} (avg {avg_shift_wide:.0f}ms)")
print(f" Onset F1: {post_f1_3:.4f} -> {post_f1_wide:.4f}")
# Step 7: Global offset correction
print("\nStep 7: Correcting systematic offset...")
midi_data, offset = apply_global_offset(midi_data, ref_onsets)
print(f" Applied {offset*1000:+.1f}ms global offset")
# Step 7b: Rhythm consolidation — snap stray onsets to dominant pattern
print("\nStep 7b: Consolidating rhythm pattern...")
midi_data, rhythm_snapped, n_dominant = consolidate_rhythm(midi_data, y, sr, hop_length)
print(f" Snapped {rhythm_snapped} notes to {n_dominant} dominant subdivisions")
# Step 7c: Merge repeated consecutive same-pitch notes without real re-attack
print("\nStep 7c: Merging repeated notes without re-attack energy...")
midi_data, notes_merged = merge_repeated_notes(midi_data, y, sr, hop_length)
print(f" Merged {notes_merged} repeated notes into sustains")
# Step 8: Fix overlaps and enforce min duration (LAST — after all position changes)
print("\nStep 8: Fixing overlaps and enforcing min duration...")
midi_data, notes_trimmed, durations_enforced = fix_note_overlap(midi_data)
print(f" Trimmed {notes_trimmed} overlapping notes")
print(f" Enforced min duration on {durations_enforced} notes")
# Step 8b: CQT-based duration extension
print("\nStep 8b: Extending note durations to match audio decay...")
midi_data, notes_extended = extend_note_durations(midi_data, y, sr, hop_length)
print(f" Extended {notes_extended} notes to match audio CQT decay")
# Step 8c: Re-enforce minimum duration after CQT extension
min_dur_enforced_2 = 0
for instrument in midi_data.instruments:
for note in instrument.notes:
if note.end - note.start < 0.10:
note.end = note.start + 0.10
min_dur_enforced_2 += 1
if min_dur_enforced_2:
print(f"\nStep 8c: Re-enforced min duration on {min_dur_enforced_2} notes after CQT extension")
# Step 8d: CQT pitch-specific energy filter (remove bass hallucinations)
print("\nStep 8d: Removing pitch-unconfirmed bass notes...")
midi_data, cqt_removed = remove_pitch_unconfirmed_notes(midi_data, y, sr, hop_length)
print(f" Removed {cqt_removed} notes with no CQT energy at their pitch")
# Step 8e: Recover missing notes from CQT energy
# Runs late so the coverage map reflects what actually survived all filters.
# Recovered notes won't be touched by phantom/spurious/pitch filters.
print("\nStep 8e: Recovering missing notes from CQT analysis...")
# Collect existing onset times to snap recovered notes to
existing_onsets = sorted(set(
round(n.start, 4) for inst in midi_data.instruments for n in inst.notes
))
midi_data, notes_recovered = recover_missing_notes(
midi_data, y, sr, hop_length, snap_onsets=existing_onsets
)
print(f" Recovered {notes_recovered} notes from CQT energy")
# Step 8f: Remove hand outliers — notes too far from their hand's cluster
print("\nStep 8f: Removing hand outlier harmonics...")
midi_data, outliers_removed = remove_hand_outliers(midi_data)
print(f" Removed {outliers_removed} outlier notes")
# Step 8f2: Enforce hand span — no chord wider than an octave per hand
print("\nStep 8f2: Enforcing hand span limit (max 12 semitones per hand)...")
midi_data, span_removed = enforce_hand_span(midi_data, max_span=12)
print(f" Removed {span_removed} notes exceeding hand span")
# Step 8g: Playability filter — limit per-onset chord size
# Complex pieces get 5 notes/hand to preserve dense voicings
# Left hand (bass) gets a tighter limit to avoid muddy chords
max_rh = 3 if complexity == 'complex' else 2
max_lh = 2 if complexity == 'complex' else 1
print(f"\nStep 8g: Playability filter (RH max {max_rh}, LH max {max_lh} per chord)...")
midi_data, playability_removed = limit_concurrent_notes(
midi_data, max_per_hand=max_rh, max_left_hand=max_lh
)
print(f" Removed {playability_removed} excess chord notes")
# Step 8h: Limit total concurrent sounding notes
print(f"\nStep 8h: Concurrent sounding limit (RH max {max_rh}, LH max {max_lh})...")
midi_data, sustain_trimmed = limit_total_concurrent(
midi_data, max_per_hand=max_rh, max_left_hand=max_lh
)
print(f" Trimmed {sustain_trimmed} sustained notes to reduce pileup")
# Step 9: Final rhythm consolidation — re-snap after all note manipulation
# Steps 8b-8h may have shifted notes off the grid. This pass catches stragglers.
# Uses wider snap (100ms) to aggressively force notes onto the grid.
print("\nStep 9: Final rhythm consolidation...")
midi_data, rhythm_snapped_2, n_dominant_2 = consolidate_rhythm(
midi_data, y, sr, hop_length, max_snap=0.10
)
print(f" Re-snapped {rhythm_snapped_2} notes to {n_dominant_2} dominant subdivisions")
# Final metrics
final_onsets = []
for inst in midi_data.instruments:
for n in inst.notes:
final_onsets.append(n.start)
final_onsets = np.unique(np.round(np.sort(final_onsets), 4))
final_f1 = onset_f1(ref_onsets, final_onsets)
final_notes = sum(len(inst.notes) for inst in midi_data.instruments)
# Duration sanity check
all_durs = [n.end - n.start for inst in midi_data.instruments for n in inst.notes]
min_dur = min(all_durs) * 1000 if all_durs else 0
print(f"\nDone:")
print(f" Phantoms removed: {phantoms_removed}")
print(f" Pitch ceiling removed: {ceiling_removed}")
print(f" Playability filter: {playability_removed} chord / {sustain_trimmed} sustain")
print(f" Chords aligned: {chords_aligned}")
print(f" Notes quantized: {notes_quantized} ({detected_tempo:.0f} BPM)")
print(f" Onsets corrected: {corrections}/{n_chords}")
print(f" Spurious onsets removed: {spurious_onsets} ({spurious_notes} notes)")
print(f" FN recovery corrections: {corrections_wide}")
print(f" Global offset: {offset*1000:+.1f}ms")
print(f" Overlaps trimmed: {notes_trimmed}")
print(f" Min durations enforced: {durations_enforced}")
print(f" Notes extended (CQT decay): {notes_extended}")
# Playability check: max concurrent notes per hand
all_final = sorted(
[n for inst in midi_data.instruments for n in inst.notes],
key=lambda n: n.start
)
max_left = 0
max_right = 0
for i, note in enumerate(all_final):
is_right = note.pitch >= 60
concurrent = sum(1 for o in all_final
if o.start <= note.start < o.end
and (o.pitch >= 60) == is_right)
if is_right:
max_right = max(max_right, concurrent)
else:
max_left = max(max_left, concurrent)
print(f" Final onset F1: {final_f1:.4f}")
print(f" Min note duration: {min_dur:.0f}ms")
print(f" Max concurrent: L={max_left} R={max_right}")
print(f" Notes: {total_notes} -> {final_notes}")
# Final step: shift all notes so music starts at t=0
# (must be AFTER all audio-aligned processing like onset detection, CQT filters)
if music_start > 0.1:
print(f"\nShifting all notes by -{music_start:.2f}s so music starts at t=0...")
for instrument in midi_data.instruments:
for note in instrument.notes:
note.start = max(0, note.start - music_start)
note.end = max(note.start + 0.01, note.end - music_start)
midi_data.write(str(output_path))
print(f" Written to {output_path}")
# Step 9: Spectral fidelity analysis (CQT comparison)
print("\nStep 9: Spectral fidelity analysis (CQT comparison)...")
try:
from spectral import spectral_fidelity
spec_results = spectral_fidelity(y, sr, midi_data, hop_length)
print(f" Spectral F1: {spec_results['spectral_f1']:.4f}")
print(f" Spectral Precision: {spec_results['spectral_precision']:.4f}")
print(f" Spectral Recall: {spec_results['spectral_recall']:.4f}")
print(f" Spectral Similarity: {spec_results['spectral_similarity']:.4f}")
# Save spectral report alongside MIDI
import json
report_path = str(output_path).replace('.mid', '_spectral.json')
Path(report_path).write_text(json.dumps(spec_results, indent=2))
print(f" Report saved to {report_path}")
except Exception as e:
print(f" Spectral analysis failed: {e}")
# Step 10: Chord detection
print("\nStep 10: Detecting chords...")
try:
from chords import detect_chords
chords_json_path = str(Path(output_path).with_name(
Path(output_path).stem + "_chords.json"
))
chord_events = detect_chords(str(output_path), chords_json_path)
print(f" Detected {len(chord_events)} chord regions")
except Exception as e:
print(f" Chord detection failed: {e}")
chord_events = []
return midi_data
def onset_f1(ref_onsets, est_onsets, tolerance=0.05):
"""Compute onset detection F1 score."""
if len(ref_onsets) == 0 and len(est_onsets) == 0:
return 1.0
if len(ref_onsets) == 0 or len(est_onsets) == 0:
return 0.0
matched_ref = set()
matched_est = set()
for i, r in enumerate(ref_onsets):
diffs = np.abs(est_onsets - r)
best = np.argmin(diffs)
if diffs[best] <= tolerance and best not in matched_est:
matched_ref.add(i)
matched_est.add(best)
precision = len(matched_est) / len(est_onsets) if len(est_onsets) > 0 else 0
recall = len(matched_ref) / len(ref_onsets) if len(ref_onsets) > 0 else 0
if precision + recall == 0:
return 0.0
return 2 * precision * recall / (precision + recall)
if __name__ == "__main__":
import sys
if len(sys.argv) < 3:
print("Usage: python optimize.py <original_audio> <midi_file> [output.mid]")
sys.exit(1)
audio_path = sys.argv[1]
midi_path = sys.argv[2]
out_path = sys.argv[3] if len(sys.argv) > 3 else None
optimize(audio_path, midi_path, out_path)