Skip to content

CHANGES-FOR-PR-2.md

Latest commit

 

History

History
112 lines (74 loc) · 4.38 KB

CHANGES-FOR-PR-2.md

File metadata and controls

112 lines (74 loc) · 4.38 KB

Voice-Pro Critical Fixes — Wave 2

Target Repository: https://github.com/abus-aikorea/voice-pro
Date: 2026-04-28
Status: Ready for PR review
Based on: Real-world testing on Ubuntu 24.04.4 minimal install

1. What This Wave Fixes

This wave addresses installation and runtime failures discovered when testing the clean fixed version on a fresh Ubuntu 24.04.4 system. These are follow-up fixes to the first wave (CHANGES-FOR-PR.md).

2. Fixes

Fix 10: Missing cmake in System Dependencies (configure.sh)

Problem: On fresh Ubuntu minimal installs (including 24.04.4), cmake is not installed by default. Several pip packages require compilation via CMake during install, causing the build to fail with:

CMake Error: CMake was unable to find a build program corresponding to "Unix Makefiles".
CMAKE_MAKE_PROGRAM is not set.  You probably need to select a different build tool.

Changes:

  • Added cmake to the apt-get install line in configure.sh alongside git, ffmpeg, and build-essential

Files changed: configure.sh

Fix 11: ctranslate2 cuDNN Fix — Shell Quoting Bug (one_click.py)

Problem: The post-install cuDNN 8 library fix in one_click.py used an inline python -c "..." string with escaped quotes (\"ctranslate2.libs\"). When passed through subprocess.run(cmd, shell=True), the shell interpreted the inner quotes and mangled the Python syntax, causing:

SyntaxError: invalid syntax
libs = os.path.join(sp, ctranslate2.libs) if sp else ;

Changes:

  • Replaced the inline python -c shell command with a temporary Python script file (/tmp/fix_cudnn8.py)
  • The script is written to disk and executed directly, completely avoiding shell quoting issues

Files changed: one_click.py

Fix 12: torchaudio.AudioMetaData Missing (start-voice.py)

Problem: pyannote.audio (a dependency of whisperx) expects torchaudio.AudioMetaData to be available at the top level (torchaudio.AudioMetaData). In some torchaudio builds — especially CPU wheels or certain CUDA configurations — this class is only accessible via torchaudio.backend.common.AudioMetaData and is not re-exported at the top level. This causes:

AttributeError: module 'torchaudio' has no attribute 'AudioMetaData'

Changes:

  • Added a defensive monkey-patch in start-voice.py (alongside the existing Gradio 5.x patches)
  • If torchaudio.AudioMetaData is missing, imports it from torchaudio.backend.common and attaches it
  • Wrapped in try/except so it never crashes if the import path changes in future versions

Files changed: start-voice.py

Fix 13: Outdated yt-dlp Pin (requirements-voice-*.txt)

Problem: Both requirements-voice-gpu.txt and requirements-voice-cpu.txt pinned yt-dlp==2025.11.12. This version is >5 months old and fails on modern YouTube due to SABR streaming and JS challenge changes.

Changes:

  • Updated to yt-dlp>=2026.3.17 in both requirements files

Files changed: requirements-voice-gpu.txt, requirements-voice-cpu.txt

3. Files Changed (Wave 2)

File Fix # Description
configure.sh 10 Add cmake to apt dependencies
one_click.py 11 ctranslate2 cuDNN fix → temp script (avoids shell quoting)
start-voice.py 12 Monkey-patch torchaudio.AudioMetaData for pyannote.audio
requirements-voice-gpu.txt 13 yt-dlp>=2026.3.17
requirements-voice-cpu.txt 13 yt-dlp>=2026.3.17

Total: 5 files changed, 0 files removed.

4. Combined Impact (Wave 1 + Wave 2)

Metric Wave 1 Wave 2 Total
Fixes 9 4 13
Files changed 24 5 29
Issues addressed #76, #62, #60 — 3 upstream issues

5. Testing Environment

  • OS: Ubuntu 24.04.4 LTS (minimal install)
  • Test path: ./configure.sh → ./start.sh
  • Result: All fixes verified — install completes, app launches, YouTube downloader works

6. Suggested PR Title for Wave 2

fix: cmake dependency, ctranslate2 shell quoting, torchaudio.AudioMetaData, and yt-dlp version

Follow-up fixes discovered during Ubuntu 24.04.4 testing: adds cmake to configure.sh, replaces inline shell script with temp file for ctranslate2 cuDNN fix, patches missing torchaudio.AudioMetaData for pyannote.audio compatibility, and updates yt-dlp pin to 2026.3.17+.