Spaces:
Sleeping
Sleeping
Commit History
-
Winner-takes-all drum transcription + harsher piano rhythm grid
Ewan +1: Claude Opus 4.6Drums: NMF components all fire at every onset (broadband), so switched from independent peak-picking per component to winner-takes-all — detect onsets once, assign each to the component with highest activation. Only allows kick+cymbal secondary when >60% of primary. Cymbal sub-classified by decay envelope (hihat/crash/ride). Piano: cap dominant rhythm bins at 4 (was 6), raise merge threshold to 2x median (was 1.5x), reduce RH concurrent to 3/2 complex/normal (was 4/3), increase final consolidation snap to 100ms. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>8b89f6d -
NMF-based drum transcription, wider final rhythm snap
Ewan +1: Claude Opus 4.6Replace spectral threshold drum classification with NMF decomposition. Decomposes the magnitude spectrogram into 3 frequency-sorted components (kick/snare/cymbal) using sklearn NMF via librosa.decompose. Extracts onsets from each activation curve using half-wave rectified first difference with peak picking. Cymbal onsets sub-classified into hihat/crash/ride by decay envelope analysis. This is the standard MIR approach (Wu & Lerch 2015) and fundamentally more robust than per-onset spectral feature classification. Piano: increase final rhythm consolidation snap from 60ms to 80ms for more aggressive cleanup of stray timing after playability filters. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>15d315f -
Fix drum classification and tighten piano rhythm grid
Ewan +1: Claude Opus 4.6Drums: rewrite classification as priority decision tree instead of multi-label independent detection. Key changes: - Kick requires sub_bass_r > 25% (was low_r > 35%), much less false positives - Snare distinguished from toms by mid_flatness > 0.04 (noise from wires) - Toms get proper detection path (low flatness + mid frequency = tonal) - Simultaneous kick+cymbal only when both bands are clearly independent - Single-path returns prevent over-classification Piano rhythm: cap dominant bins at 6 (was 8), raise min_count to 4%/3. Add 8th-note fallback grid so notes too far from dominant positions still get quantized rather than being left unsnapped. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>cb9bb7e -
Real drum samples from LinnDrum CDN, tighter rhythm and fewer notes
Ewan +1: Claude Opus 4.6Replace synthesized drum samples with real LinnDrum (LM-2) recordings from smpldsnds CDN (OGG format, CORS-enabled). Removes 412KB base64 module, bundle drops from 919KB to 497KB. Piano rhythm: rewrite consolidate_rhythm to pick top bins by count (capped at 8 max) instead of threshold-based selection. Increase max_snap to 60ms. Reduce concurrent note limits (RH: 4->3, LH: 2->1 for normal; 5->4, 3->2 for complex). Raise merge_repeated_notes energy threshold from 1.2x to 1.5x median so more re-attacks merge into sustains. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>f8777bb -
Sample-based drum kit and tighter rhythm snapping
Ewan +1: Claude Opus 4.6Replace synthesized drums with DSP-generated WAV samples encoded as base64 in a JS module (no binary files). Kick/snare/hihat/crash/ride/toms use proper filtered synthesis (scipy bandpass, layered noise+sine) rendered offline for much better sound quality than real-time Tone.js synthesis. Loaded via Tone.Players from data URIs. Piano: tighten rhythm consolidation - dominant bin threshold raised to 25%/5%, add second consolidation pass (Step 9) after all note manipulation with 60ms snap range to catch notes shifted off-grid by CQT extension and playability filters. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>8fbe43c -
Tighten rhythm snapping and fix drum sync
Ewan +1: Claude Opus 4.6Piano: increase consolidate_rhythm max_snap from 40ms to 50ms, tighten dominant bin threshold from 15%/3% to 25%/5% so fewer subdivisions qualify and notes snap to a cleaner rhythmic grid. Drums: add global offset correction step that measures median timing drift between drum events and audio onsets, then shifts all events to compensate. This fixes drums being out of sync with the original audio overlay. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>84b4faf
-
LH anchors on melody (top note), enforce span on concurrent notes too
Ewan +1: Claude Opus 4.6Fundamental voicing philosophy change: both hands now protect the MELODY (highest note) and build harmonics downward. Previously LH protected the bass (lowest note), which led to wide unplayable voicings. Changes across three functions: - enforce_hand_span: Now checks concurrent sounding notes (sustained overlaps), not just chord groups. Both hands anchor on highest note. Sustained notes exceeding span get trimmed rather than removed. - limit_concurrent_notes: LH now protects highest (melody) not lowest - limit_total_concurrent: Same - LH protects melody voice - remove_hand_outliers: Both hands flag lowest outlier, protecting melody Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>9c5db09 -
Enforce hand span limit — no chord wider than an octave per hand
Ewan +1: Claude Opus 4.6New enforce_hand_span() step based on empirical pianist hand span data (Boyle et al. 2015, n=473): average adult comfortable span is 12 semitones (one octave), with 14-16 being a stretch for large hands. For each chord, if any hand spans more than 12 semitones: - RH: protects melody (highest note), removes lowest until span fits - LH: protects bass (lowest note), removes highest until span fits This catches physically impossible voicings that passed the note-count limiter — e.g. 2 notes spanning 15 semitones is within the count limit but unplayable for most pianists. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>5c1af8d -
Merge repeated same-pitch notes into sustains, increase drum detection sensitivity
Ewan +1: Claude Opus 4.6Piano: New merge_repeated_notes() step (7c) checks onset energy at re-attack points of consecutive same-pitch notes. If re-attack energy is below 1.2x the median onset strength, the notes are merged into one sustained note instead of stuttering re-strikes. Preserves real repeated notes with genuine attack energy. Drums: Significantly increased detection sensitivity: - Hi-hat band: delta 0.05->0.03, RMS threshold 0.003->0.001, wait 3->2 - Mid band: delta 0.06->0.04, RMS 0.005->0.003, wait 3->2 - Low band: delta 0.08->0.06 - Added full-band safety net pass (catches hits all sub-bands miss) - Improved fallback classifier for full-band-only detections - Merge window 25ms->30ms for better cross-band alignment Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>f474670 -
Add rhythm consolidation step to discipline note onsets
Ewan +1: Claude Opus 4.6After onset correction pulls notes toward audio onsets, stray timings can scatter across many micro-positions, losing rhythmic clarity. New consolidate_rhythm() step: 1. Builds histogram of note positions within each beat (16 bins = 16th notes) 2. Identifies dominant subdivisions (bins with significant note density) 3. Re-snaps stray onsets to the nearest dominant position (within 40ms) Notes already on dominant positions are untouched. This preserves the melodic contour while disciplining timing to the song's actual rhythmic pattern. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>1f67352 -
Remove hand outlier harmonics, rewrite drum transcription with multi-band detection
Ewan +1: Claude Opus 4.6Piano: Add remove_hand_outliers step that detects notes isolated from their hand cluster (>7 semitones from nearest neighbor). Catches high harmonics in LH and low harmonics in RH that cannot be physically played. Drums: Rewrite transcription with sub-band onset detection - filters into low (<250Hz), mid (200-5000Hz), high (>3000Hz) bands and detects onsets independently. Classification based on which bands triggered (e.g. snare = mid+high, kick = low-only, hihat = high-only). Handles simultaneous hits naturally. Uses 5-band spectral energy aligned to actual drum physics. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>79912b8 -
Use solo piano optimizer for full-song melodic stem, reduce LH bass notes
Ewan +1: Claude Opus 4.6Switch full-song pipeline from optimize_other to the full solo piano optimizer for the melodic stem — produces much better rhythm, playability, and note accuracy. Also reduces left-hand concurrent notes limit (2 normal, 3 complex) to avoid muddy bass chords while keeping right hand at 4-5. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>aa08171
-
Improve transcription fidelity: trailing notes, sustain pedal, complexity tuning
Ewan +1: Claude Opus 4.6- Fix trailing silence threshold (5% → 2% RMS + 3s protection zone) — recovers 7s of cut-off endings - Fix leading silence threshold (10% → 5% + always protect first note) - Add spectral masking in harmonic ghost removal for two-hand texture - Add sustain pedal detection from audio spectral flux analysis - Add complexity-aware tuning (note density + polyphony estimation) - Add audio analysis toolkit (spectral comparison, CQT visualization) - UI: larger transport icons, loop label, 5s skip labels Jewish Bride spectral MSE: -40.3% overall, -73.8% at 95th percentile Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>1646c97 -
Enhance spurious onset filter for higher precision
Ewan +1: Claude Opus 4.6- Lower Cat 5 distance threshold from 200ms to 120ms (catches more hallucinated multi-note onsets far from audio) - Add Cat 6: remove onsets where all notes are <50ms (splinter artifacts) - Add Cat 7: remove onsets >70ms from audio with strength <2.5 (near-miss hallucinations just outside matching window) - Benchmarks: F1 improved across all songs (Adele 0.929→0.953, Beatles 0.963→0.974, IDWBYA 0.918→0.952, WTPO 0.963→0.971, Jorja 0.948→0.952) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>7dd6b8a -
Improve register balance: CQT-check bass, relax upper threshold
Ewan +1: Claude Opus 4.6- Extend pitch-unconfirmed CQT filter to check MIDI 40-47 (bass) with -35dB threshold, catching bass notes without real spectral energy - Relax upper register threshold from -20dB to -25dB, preserving more legitimate notes above C5 - Results: bass notes reduced ~1-2%, upper notes increased 5-40% across benchmarks with F1 maintained or improved Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>81f7585 -
Fix tempo doubling and reduce harmonic ghost over-removal
Ewan +1: Claude Opus 4.6- Halve detected tempo when >140 BPM and half-tempo is reasonable (50-120), fixing songs like Billie Eilish IDWBYA being detected at 172 instead of 86 - Tighten harmonic ghost velocity ratio thresholds (0.75→0.55 general/C5+, 0.95→0.85 sub-bass) to preserve more legitimate notes on clean piano - Benchmarked: IDWBYA F1 improved 0.905→0.917, Adele retains 103 more notes Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>307ea81 -
Initial commit - Mr Octopus piano tutorial app
Ewanf0a176a