Spaces:
Sleeping
Sleeping
Commit History
-
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 -
Fix MetalSynth API (HH/ride/crash) and use mix audio for beat tracking
Ewan +1: Claude Opus 4.6- MetalSynth.triggerAttackRelease in Tone.js v14+ requires frequency as first param: (freq, duration, time, velocity). Old code used r13 sig (duration, time, velocity) which silently failed. Now passes explicit Hz values for hihat (400), crash (300), ride (500). - Use original mixed audio for beat tracking in optimize_other instead of the separated stem. The full mix has drums/bass which drive the beat, giving much more accurate tempo detection and beat grid. Onset detection still uses the clean separated stem. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>ecae4b3 -
Fix piano roll quality, drum sounds, add original audio playback, fix speed
Ewan +1: Claude Opus 4.6- Add playability limits to full-song optimizer (limit_concurrent_notes, limit_total_concurrent, remove_harmonic_ghosts) — reduces note count to physically playable levels - Fix drum hi-hat not audible: increase MetalSynth volume -12 -> -4dB, extend trigger duration 32n -> 16n, boost crash/ride volumes too - Add original audio playback: new /api/jobs/{id}/audio endpoint serves uploaded file back, HTML Audio element synced to Transport with toggle button + volume slider in Controls - Fix speed scaler: preserve playbackRate via tempoRef across scheduleNotes reschedules (transport.stop() was losing the rate) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>b5c16f8 -
Add drum transcription with madmom onset detection + spectral classification
Ewan +1: Claude Opus 4.6New feature: Drums tab in full song mode. Uses madmom RNNOnsetProcessor for high-quality onset detection on the Demucs drums stem, then classifies each hit (kick/snare/hi-hat/crash/ride/tom) by spectral content using librosa features (centroid, flatness, ZCR, frequency band energy). - transcriber/drums.py: drum transcription pipeline - DrumSheet.jsx: lane-based drum grid renderer (color-coded by instrument, diamonds for cymbals, X for hi-hat, circles for kick/snare/toms) - usePlayback.js: synthesized drum kit (MembraneSynth for kick/toms, NoiseSynth+MembraneSynth for snare, MetalSynth for hi-hat/crash/ride) - Dockerfile: install madmom from git (PyPI version incompatible with 3.12) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>eed42a3
-
Fix bass in piano roll, upgrade guitar samples, improve bass optimizer
Ewan +1: Claude Opus 4.6- Piano roll: use pianoRollNotes (bass filtered) for audio scheduling, not full notes array. Bass was bleeding through to piano playback. - Guitar samples: switch from Magenta SGM+ (sparse MP3 every 3 semitones) to tonejs-instruments (University of Iowa chromatic OGG recordings for acoustic, Karoryfer for electric). Much warmer, less tinny. - Bass optimizer: add pitch floor (E1=28), percussion bleed removal (short notes + spectral centroid check), octave error correction (contour-based), isolated note removal. Should produce cleaner bass lines with fewer wrong notes. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>1c51e5f
-
Remove unused vocal transcription and arrangement from full song pipeline
Ewan +1: Claude Opus 4.6The arrangement engine output is no longer used (piano roll shows the "other" stem directly). Skip vocal transcription, vocal optimization, and arrangement generation to save ~15-20s of processing time. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>477c19c -
Use direct melodic transcription for piano roll instead of arrangement
Ewan +1: Claude Opus 4.6The arrangement engine (vocals + chord voicings) was producing poor results because vocal transcription via Basic Pitch is inherently noisy and chord detection on mixed stems compounds errors. Now the piano roll uses the "other" stem transcription directly — the same data that produces good guitar tabs. This is the actual transcribed melodic instruments (guitar, piano, synths) rather than a lossy derivation. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>a14f449 -
Improve transcription quality + upgrade to Magenta SGM+ samples
Ewan +1: Claude Opus 4.6Bass optimizer: - Lower pitch ceiling from MIDI 64 to 55 (G3, full 4-string bass range) - Add monophonic enforcement (keep lowest pitch at each onset) - Add spurious onset removal - Fix overlap trimming to trim against ANY next note (bass is monophonic) Vocals optimizer: - Add pitch ceiling at MIDI 84 (C6, soprano range) - Add pitch floor at MIDI 40 (E2, below = bass territory) - Add phantom high note removal - Add spurious onset removal Samples: Replace tonejs-instruments with Google Magenta SGM+ soundfont (acoustic_guitar_steel, electric_guitar_clean, electric_bass_finger). Higher quality source recordings, professional soundfont. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>83ff7ad -
Sample-based guitar/bass sounds, filter bass from piano roll, two guitar tabs
Ewan +1: Claude Opus 4.6Replace synthesized PluckSynth/MonoSynth with real recorded samples from tonejs-instruments (acoustic guitar, electric guitar, bass electric). Filter bass notes from piano roll in full song mode. Split guitar tab into Acoustic Guitar and Electric Guitar views. Update timing estimate. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>aeb0fce -
Full Song quality pass: light optimizer, better guitar sound, tab accuracy
Ewan +1: Claude Opus 4.6- New optimize_other.py: skips 7 piano-specific steps (CQT recovery, duration extension, harmonic ghost removal, playability limits) that corrupt mixed-instrument content from the Demucs "other" stem - Guitar synth: Karplus-Strong PluckSynth pool with chorus + reverb replaces synthetic FM triangle oscillator - Bass synth: MonoSynth with filter envelope (finger-pluck brightness decay) + compressor replaces basic triangle wave - Tab accuracy: polyphony cap (6 notes), fret span constraint (5 frets), duration cap (2s), EMA hand position, tighter pitch range (40-84) - Arrangement: sparser RH fill (1s gap threshold, 2 tones, vel 45) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>78ef95d -
Add piano arrangement engine + guitar/bass tab views for Full Song mode
Ewan +1: Claude Opus 4.6Full Song mode now generates a proper piano arrangement (LH chord voicings, RH vocal melody with harmonic fill) instead of raw stem dumps. Adds scrolling canvas-based guitar and bass tab views with playback tracking. New backend pipeline: optimize_vocals → arrange_piano → tabs generation. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>1a538e9 -
Add Full Song mode with Demucs source separation
Ewan +1: Claude Opus 4.6New "Full Song" tab on the upload screen uses Demucs to separate any audio into stems, then transcribes piano and bass parts independently. Bass notes render in amber on the piano roll. - New transcriber/separate.py: Demucs wrapper - New transcriber/optimize_bass.py: simplified bass optimization - New API endpoint POST /api/transcribe-full with async polling - Upload screen: Solo Piano / Full Song tab switcher - Color scheme: bass instrument color (amber) - MIDI parser: instrument detection from MIDI program number - Dockerfile: CPU-only PyTorch + Demucs + pre-downloaded model Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>56c8033
-
Initial commit - Mr Octopus piano tutorial app
Ewanf0a176a