Skip to content

CHANGES-FOR-PR.md

Latest commit

 

History

History
178 lines (122 loc) · 8.67 KB

CHANGES-FOR-PR.md

File metadata and controls

178 lines (122 loc) · 8.67 KB
Raw

Voice-Pro Critical Fixes — Pull Request Summary

Target Repository: https://github.com/abus-aikorea/voice-pro
Date: 2026-04-27
Status: Ready for PR review

1. What This PR Fixes

This PR addresses runtime failures that prevent Voice-Pro from being operational on modern systems. All changes are backward-compatible and additive — the original ./start.sh launch path works exactly as before.

Upstream Issues Addressed

Issue Title Fix # Note
#76 none_type not iterable error 3 Gradio 5.x Progress.tqdm() returns self instead of iterable; patched in start-voice.py
#62 Installation Error 4 setuptools>=80 removed pkg_resources; one_click.py now pins <70 and installs whisper without build isolation
#60 Installation is too much trouble 4, 6 Installer reliability improved; GPU auto-detect removes manual config step

Issues not addressed by this PR: #61 (Azure SDK TranslatorCredential), #75/#74 (RTX 5070 sm_120 ctranslate2 kernel), #79 (azure-ai-translation-text 1.0.0+), #80 (demucs subprocess failure detection), #68 (missing demucs output files), #67 (Windows grep compatibility).

2. Critical Fixes

Fix 1: YouTube Downloader (app/abus_downloader.py)

Problem: The built-in YouTube downloader fails on most videos because:

  • yt-dlp 2025.11.12 is >5 months old and cannot handle YouTube's current SABR streaming + JS challenges
  • No JavaScript runtime is configured (yt-dlp 2026+ only enables Deno by default)
  • Metadata is extracted twice per download, doubling HTTP requests and triggering rate limits

Changes:

  • Updated ydl_opts with js_runtimes: {'node': {}} and remote_components: ['ejs:github']
  • Replaced double extraction with single extract_info(..., download=False) + process_ie_result(info, download=True)

Note: yt-dlp itself must also be updated to 2026.3.17+ in the environment (done via pip in the conda env).

Fix 2: VAD Model Path (src/vad.py)

Problem: The VAD constructor hardcodes ~/.cache/whisper-live/silero_vad.onnx and ignores the project-local copy at model/vad/silero_vad.onnx. On fresh clones or systems without the cache, VAD fails immediately.

Changes:

  • Constructor now checks model/vad/silero_vad.onnx (project-local) first
  • Falls back to ~/.cache/whisper-live/silero_vad.onnx
  • Prints clear manual-sourcing instructions if neither is found

Fix 3: Gradio 5.x Compatibility (app/abus_path.py, app/gradio_*.py, start-voice.py)

Problem: Gradio 5.x introduced breaking API changes:

  • gr.Progress.tqdm() returns self instead of a wrapped iterable, causing "NoneType object is not iterable" errors (reported in upstream issue #76)
  • gr.File(type='filepath') now returns a NamedString (str subclass with .name) instead of a plain str, while gr.Audio(type='filepath') still returns a plain str. Code using file_obj.name crashes when the input is a plain str.
  • FileData.model_validate() rejects payloads with missing/None meta fields, causing upload failures.

Changes:

  • Added gradio_file_path(file_obj) helper in app/abus_path.py that handles str, NamedString, and legacy file objects
  • Updated 13 app/gradio_*.py files to use gradio_file_path(file_obj) instead of file_obj.name
  • Added monkey-patches in start-voice.py for Progress.tqdm (wraps with standard tqdm) and FileData.model_validate (injects default meta when missing)

Files affected: app/abus_path.py, app/gradio_aicover.py, app/gradio_asr.py, app/gradio_demixing.py, app/gradio_gulliver.py, app/gradio_kara.py, app/gradio_rvc.py, app/gradio_translate.py, app/gradio_tts_cosyvoice.py, app/gradio_tts_edge.py, app/gradio_tts_f5.py, app/gradio_tts_kokoro.py, app/gradio_tts_rvc.py, app/gradio_vsr.py, start-voice.py

Fix 4: Installer Build Fixes (one_click.py)

Problem: Fresh installs fail on modern Python environments because:

  • setuptools>=80 removed pkg_resources, breaking openai-whisper (sdist that imports pkg_resources during build)
  • openai-whisper build isolation pulls in a newer setuptools that lacks pkg_resources
  • ctranslate2 bundles a renamed cuDNN 8 library but omits the split .so files (e.g., libcudnn_ops_infer.so.8) that it dlopens at runtime, causing CUDA errors
  • The inline python -c shell command for the ctranslate2 fix had quoting issues that caused SyntaxError on some shells

Changes:

  • Pre-install setuptools<70 and wheel before requirements
  • Install openai-whisper==20240930 with --no-build-isolation
  • Post-install: write the cuDNN fix to a temp Python script file instead of an inline python -c string, avoiding shell quoting bugs

Fix 5: cuDNN Version Bump (requirements-voice-gpu.txt)

Problem: nvidia-cudnn-cu12==8.9.7.29 pins an old version that may conflict with newer PyTorch builds.

Changes:

  • Relaxed to nvidia-cudnn-cu12>=9.1.0.70

Fix 6: GPU Auto-Detection (start.sh, update.sh)

Problem: Users on fresh systems often forget to set GPU_CHOICE, causing the app to default to an incorrect mode or prompt interactively.

Changes:

  • Added auto-detection: if nvidia-smi is available, set GPU_CHOICE=G; otherwise GPU_CHOICE=C

Fix 7: Build Tool Dependencies (configure.sh)

Problem: Ubuntu minimal installs (and some fresh 24.04 setups) lack cmake, causing pip packages that require compilation to fail with CMAKE_MAKE_PROGRAM is not set.

Changes:

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

Fix 8: torchaudio / pyannote.audio Compatibility (start-voice.py)

Problem: pyannote.audio (dependency of whisperx) expects torchaudio.AudioMetaData to be available at the top level. In some torchaudio builds (especially CPU wheels or certain CUDA builds), this class is not re-exported, causing AttributeError: module 'torchaudio' has no attribute 'AudioMetaData' on import.

Changes:

  • Added monkey-patch in start-voice.py: if torchaudio.AudioMetaData is missing, import it from torchaudio.backend.common and attach it

Fix 9: yt-dlp Version Pin (requirements-voice-*.txt)

Problem: yt-dlp==2025.11.12 is too old for modern YouTube.

Changes:

  • Relaxed to yt-dlp>=2026.3.17 in both GPU and CPU requirements files

3. Files Changed

File Fix # Description
app/abus_downloader.py 1 YouTube JS runtime + single extraction
src/vad.py 2 Local VAD path check first
app/abus_path.py 3 gradio_file_path() helper
app/gradio_aicover.py 3 Use gradio_file_path()
app/gradio_asr.py 3 Use gradio_file_path()
app/gradio_demixing.py 3 Use gradio_file_path()
app/gradio_gulliver.py 3 Use gradio_file_path()
app/gradio_kara.py 3 Use gradio_file_path()
app/gradio_rvc.py 3 Use gradio_file_path()
app/gradio_translate.py 3 Use gradio_file_path()
app/gradio_tts_cosyvoice.py 3 Use gradio_file_path()
app/gradio_tts_edge.py 3 Use gradio_file_path()
app/gradio_tts_f5.py 3 Use gradio_file_path()
app/gradio_tts_kokoro.py 3 Use gradio_file_path()
app/gradio_tts_rvc.py 3 Use gradio_file_path()
app/gradio_vsr.py 3 Use gradio_file_path()
start-voice.py 3, 8 Monkey-patches for Gradio 5.x tqdm + FileData + torchaudio.AudioMetaData
one_click.py 4 setuptools, whisper, ctranslate2 cuDNN fixes (temp script)
start.sh 6 GPU auto-detect
update.sh 6 GPU auto-detect
configure.sh 7 Add cmake to apt dependencies
requirements-voice-gpu.txt 5, 9 cuDNN version bump + yt-dlp update
requirements-voice-cpu.txt 9 yt-dlp update

Total: 24 files changed, 0 files removed.

4. Testing

Test Result
YouTube download (modern video with JS challenges) ✅ Success
VAD initialization (fresh clone, no cache) ✅ Success
Gradio file upload (audio + video inputs) ✅ Success
Progress bar iteration (subtitle generation) ✅ Success
Fresh install on Linux Mint 22.2 ✅ Success

5. Suggested PR Title

fix: YouTube downloader, VAD paths, Gradio 5.x compatibility, and installer build errors

Fixes yt-dlp JS runtime config for modern YouTube, adds project-local VAD fallback, resolves Gradio 5.x breaking changes (tqdm/FileData/file paths), and patches one_click.py for setuptools/pkg_resources and ctranslate2 cuDNN compatibility.