Installation fails on Windows (CPU mode, AMD GPU) due to multiple dependency conflicts (v3.1.4 and v3.2.0)

Environment

• OS: Windows 11 (64-bit)
• GPU: AMD Radeon 780M (integrated) – running in CPU mode (C option selected)
• RAM: 32 GB
• Visual Studio Build Tools 2022 with "Desktop development with C++" installed
• Tested versions: v3.1.4 and v3.2.0 (both from the GitHub releases page)

Steps to reproduce

1. Downloaded the source code ZIP from the release page (v3.2.0 and later v3.1.4).
2. Extracted the folder to Desktop.
3. Ran configure.bat (completed successfully; selected CPU mode when prompted).
4. Ran start.bat as administrator.

Expected behavior

The web UI opens at http://127.0.0.1:7870 without errors.

Actual behavior

start.bat fails with a series of import/dependency errors. The errors occur after configure.bat finishes, and they appear every time start.bat is run, even after a clean extraction. I attempted multiple manual fixes (adjusting package versions) but each fix uncovers a new conflict. The sequence of errors encountered (across both versions) includes:

1. ModuleNotFoundError: No module named 'pkg_resources'

   • Originates from ctranslate2 (used by faster_whisper). Solved by downgrading setuptools to 67.8.0, but then other issues appear.

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

   • In pyannote.audio.core.io. Triggered by incompatibility between torchaudio and pyannote.audio. Requires specific combinations (e.g., torch==2.1.1, torchaudio==2.1.1 vs. torch==2.2.0, torchaudio==2.2.0) but each combination leads to further conflicts.

3. ImportError: huggingface-hub>=0.26.0,<1.0 is required ... but found huggingface-hub==1.13.0

   • From transformers dependency check. Downgrading huggingface-hub causes conflict with pyannote or other packages.

4. numpy.dtype size changed, may indicate binary incompatibility

   • When numpy 2.x is installed, compiled extensions (like spacy, thinc, speechbrain) crash because they were built for numpy 1.x. Installing numpy==1.26.4 resolves this but does not prevent other errors (e.g., huggingface-hub conflict returns).

5. ImportError: Lazy import of LazyModule(...speechbrain.integrations.k2_fsa...) failed

   • Occurs with speechbrain 1.1.0 and pydantic v2; downgrading to pydantic<2 causes breakage in fastapi/gradio.

6. After manually installing torch==2.2.0, torchaudio==2.2.0, pyannote.audio==3.3.2, speechbrain==1.0.1, and numpy==1.26.4, the error returned to huggingface-hub==1.13.0 being required but found.

In summary, the dependency tree appears broken on a clean Windows system when using CPU mode. configure.bat installs a mix of very recent packages (e.g., huggingface-hub 1.13.0, numpy 2.2.6, torch 2.11.0) that are incompatible with the pinned requirements of transformers, pyannote.audio, speechbrain, spacy, etc.

Additional context

• configure.bat reports success and uses Chocolatey to install Visual Studio Build Tools (which were already present).
• I tried deleting the installer_files folder and re-running configure.bat multiple times with the same result.
• Internet connection is stable; no proxy/VPN used.
• The issue persists identically on both v3.1.4 and v3.2.0.

Please advise if there is a recommended combination of package versions for CPU-only mode, or if an updated requirements-voice-cpu.txt will be released. Thank you.