SheetSage Windows Installation Feedback

SheetSage Windows Installation Feedback.md

Title

Windows local installation feedback: embedded Python 3.8 issues, broken dependency pins, missing runtime deps, asset download failures, and suggested installer improvements

---

Preamble

Hi! First of all, thank you for SheetSage — the project is very impressive and, after some work, I was able to get the full pipeline running locally on Windows (piano transcription, Demucs separation, lead sheet generation, Gradio UI).

Because the Windows setup took quite a bit of debugging, I wanted to share detailed installation feedback and concrete suggestions that may help improve the out-of-the-box experience for other Windows users.

This is not a complaint — just a consolidated report of issues encountered and fixes/workarounds that made the system work reliably on my machine.

---

Environment

• OS: Windows 10 / Windows 11
• Python: Embedded Python 3.8 (as currently used by setup scripts)
• Hardware: CPU only
• Installation method: setup_local.bat / run_local.bat

---

Summary of Issues Encountered

1. Embedded Python 3.8 missing pip / ensurepip

• The embedded Python distribution did not include pip.
• ensurepip was also missing, so pip had to be bootstrapped manually via get-pip.py (Python-3.8-specific version).

Suggestion

• Consider switching to embedded Python 3.9+ / 3.10+, or
• Automatically bootstrap pip depending on Python version.

---

2. requirements.txt pins incompatible with Python 3.8

Several pinned versions could not be installed on Python 3.8:

• attrs==25.4.0 (not available for py38)
• audioread==3.1.0 (does not exist)
• cffi==2.0.0 (no py38 wheels)
• click==8.3.1 (does not exist)

This required manually relaxing versions to get a working environment.

Suggestion

• Use compatible version ranges instead of strict pins, or
• Provide Python-version-specific constraints (e.g. constraints-win-py38.txt), or
• Officially require Python ≥3.9 and align requirements accordingly.

---

3. Runtime dependencies missing (discovered only at launch time)

Several modules were required at runtime but not installed by default:

• gradio
• torch, torchaudio, torchvision
• validators
• demucs
• piano_transcription_inference

These failures only appeared after launching the UI or starting inference.

Suggestion

• Provide extras such as:

  pip install .[ui]
  pip install .[full]

• Or add a pre-launch dependency check in run_local.bat.

---

4. FluidSynth hardcoded path on Windows

pretty_midi attempted to load FluidSynth from:

C:\tools\fluidsynth\bin

which did not exist. The project already ships FluidSynth DLLs locally, but the hardcoded path caused a crash.

Suggestion

• Remove hardcoded path and instead:

  • Use the local bin/ directory, or
  • Respect an environment variable like FLUIDSYNTH_DIR.

---

5. ffprobe / ffmpeg not found → audio loading failure

Gradio failed to load non-WAV audio files because ffprobe was not in PATH:

RuntimeError: ffprobe not found

Suggestion

• Ensure ffmpeg including ffprobe.exe is added to PATH during setup.
• Alternatively, explicitly configure pydub’s ffmpeg paths.

---

6. Asset downloader issues (assets.py)

Several problems occurred when downloading model assets:

1. Initial multiprocessing bug (args undefined / unexpected keyword argument).

2. HTTP 403 errors from:

   • https://sheetsage.s3.amazonaws.com/...

3. Fallback to yt-dlp also failed:

   • S3 URLs still returned 403
   • MEGA folder URLs are not supported by yt-dlp

Suggestion

• Fix multiprocessing argument handling (already patched locally).

• Avoid MEGA folder URLs as a primary distribution method.

• Prefer:

  • GitHub Releases with direct file links, or
  • Stable presigned S3 URLs with proper headers (User-Agent).

• If MEGA is required, use MEGAcmd rather than yt-dlp.

---

7. Gradio progress requires queue

Gradio crashed with:

ValueError: Progress tracking requires queuing to be enabled.

Fix

demo.queue()
demo.launch(...)

Suggestion

• Enable queue by default when progress updates are used.

---

8. Lead sheet formatting too strict on downbeats

Lead sheet generation initially failed with:

ValueError: Change at onset X is not on a downbeat

Snapping chord changes to the nearest downbeat instead of crashing allowed successful output.

Suggestion

• Default to snapping + warning rather than raising an exception.
• This is especially important when fallback beat tracking is used.

---

Suggested High-Level Improvement: Windows Installer

A single robust Windows installer would help significantly:

• Embedded Python ≥3.9

• Automatic pip bootstrap

• Grouped dependency installation (core, ui, optional)

• ffmpeg + fluidsynth + lilypond PATH setup

• Sanity checks before launch:

  python -c "import gradio, torch, validators"
  where ffprobe
  fluidsynth --version

---

Final Notes

After addressing the above points, SheetSage runs successfully end-to-end on Windows, including:

• Demucs vocal separation
• Piano transcription
• Lead sheet generation
• Gradio UI

I hope this feedback is useful. Thanks again for the project — it’s very promising, and a smoother Windows setup would make it much more accessible.

Happy to clarify or help test fixes if needed!