fact-lap/file-converter

File Converter

Flask/Docker file converter for images, documents, audio, and music notation.

Music / Notation Lane

Deterministic conversions:

• —.musicxml / .xml / .mxl / .mscz -> .pdf, .mid, .musicxml, .mscz, .wav, .mp3, .flac, .ogg
• —.mid / .midi -> .pdf, .musicxml, .mscz, .wav, .mp3, .flac, .ogg

Draft AI transcription:

• —audio (.mp3, .wav, .flac, .aac, .ogg, .m4a, .opus, .wma) -> .mid, .musicxml, .pdf
• —direct mode runs basic-pitch
• —demucs mode runs demucs first, then basic-pitch
• —melody cleanup can collapse noisy transcription to one monophonic line and optionally quantize to 1/8 or 1/16

The audio transcription lane is lossy and should be treated as a draft score for human cleanup.

Stem split:

• —audio -> .zip containing separate WAV stems
• —two-stem mode: vocals + instrumental
• —four-stem mode: vocals + drums + bass + other

Runtime Tools

Required system tools:

• —ffmpeg
• —musescore3 or another MuseScore CLI available as MUSESCORE_BIN
• —demucs for stem split and optional transcription pre-processing

Python dependencies are pinned in requirements.txt, including basic-pitch, demucs, music21, mido, librosa, pretty_midi, torch, and torchaudio.

MuseScore runs headless with:

QT_QPA_PLATFORM=offscreen

Verified RunPod Phase 0

Verified on RunPod pod 5k6kk9qklo23l0:

• —RTX 4090 / CUDA visible through PyTorch
• —MusicXML -> PDF
• —MusicXML -> MIDI
• —WAV -> MIDI through basic-pitch
• —WAV -> MIDI through demucs + basic-pitch
• —WAV -> draft PDF through basic-pitch + MuseScore