Research: Evaluate demucs v4 and pyguitarpro GP8 support

[carochacs]

Related SOTA Research Issues

• Pitch extraction: get-flashbacks/feedback-plugin-lyrics-karaoke#2 — CREPE accuracy depends on vocal stem quality from demucs
• Lyrics transcription: get-flashbacks/feedback-plugin-lyrics-sync#2 — WhisperX accuracy depends on vocal stem quality from demucs
• Chord detection: get-flashbacks/feedback-plugin-chord-detector#3 — Chord detection depends on good stem separation

---

Research Objective

Evaluate stem separation quality improvements in demucs v4 vs. current v3, and verify Guitar Pro format parsing supports latest GP8 versions without regressions.

Current Approach

• Stem separation: demucs v3 (Facebook Meta) + Roformer via audio-separator
• Format parsing: pyguitarpro>=0.10.1 for GP3–GP5, pure Python GPIF parser for GP6/7/8
• MusicXML: stdlib xml.etree.ElementTree (lightweight, sufficient)

SOTA Alternatives

Option 1: Demucs v4

• Pros: Improved source separation quality, better vocal/drum isolation, newer models
• Cons: Different model API, potential performance regression if slower
• Cost: Dependency bump, minor integration work

Option 2: Roformer (already in use)

• Status: Competitive with demucs v4 on many benchmarks; no urgent change needed

Option 3: UVR backend variants

• Status: Redundant given Roformer; not prioritized

Research Methodology

Phase 1: Literature Review (2–4 hours)

• Papers: Demucs v4 paper/blog post, source separation benchmarks (MUSDB18)
• GitHub: facebookresearch/demucs (branches, issues, v4 changelog)
• Benchmarks: MUSDB18 dataset, existing test songs

Phase 2: PoC Setup (4–8 hours)

1. Install demucs>=4.0.0 alongside v3
2. Export stems using both versions on test songs
3. Compare output files (bitrate, duration, quality)
4. Verify pyguitarpro on latest GP8 files (if available)

Phase 3: Benchmark (8–16 hours)

• Test corpus: 5–10 diverse songs (rock, pop, acoustic, jazz, electronic)
• Metrics:
  • Source separation quality (SDR, ISR, SIR on MUSDB18 if applicable)
  • Inference time (seconds per minute of audio, CPU vs. GPU)
  • Output bitrate / file size consistency
  • GP8 parser: verify no errors on real-world GP8 files
• A/B comparison: Manual listening, frequency-domain plots

Phase 4: Decision (2–4 hours)

• If demucs v4 SDR improvement > 1 dB AND latency acceptable → adopt
• If latency regression > 20% → stay on v3
• For GP8: if parser handles all test files → update docs, no code change needed

Success Criteria

• Benchmark dataset: 5–10 songs, diverse genres
• Primary metric: SDR (Source-to-Distortion Ratio) on MUSDB18 subset or custom eval
• Secondary metric: Inference latency (CPU/GPU time per minute audio)
• GP8 success: 100% parser pass rate on test files
• Decision rule: "Upgrade demucs v4 if SDR gain > 1 dB and latency increase < 20%"

References

• Research Plan: SOTA Review Plan
• Benchmark: MUSDB18 dataset (source separation standard)
• Format parsing: feedback/lib/gp2rs.py, feedback/lib/gp2rs_gpx.py
• Stem separation pipeline: feedback/lib/sloppak_convert.py