Replace SpeechBrain and Lightning monkey-patches with upstream upgrades

[bobsummerwill-openclaw-bot]

Summary

Replace fragile monkey-patches in install_venv.sh with upstream dependency upgrades. This removes ~160 lines of patching code from Steps 7 and 8.

Changes

Step 7: SpeechBrain (was: monkey-patch, now: upgrade)

SpeechBrain >= 1.0.3 (released March 12, 2025) includes native fixes for torchaudio 2.9.x compatibility:

• torch_audio_backend.py now wraps list_audio_backends() in try/except with fallback to torchaudio.utils.available_backends — fixes Workaround: SpeechBrain list_audio_backends() removed in torchaudio 2.9.x #38
• dataio.py now wraps torchaudio.info() in try/except with a soundfile.info() fallback — fixes Workaround: SpeechBrain torchaudio.info() removed in torchaudio 2.9.x #39

Previously we were patching these files post-install with ~120 lines of inline Python scripts.

Step 8: Lightning (was: monkey-patch, now: upgrade)

Lightning >= 2.4.0 (released mid-2024) explicitly defaults weights_only=False in cloud_io.py, maintaining backward compatibility with PyTorch 2.6+ which changed the torch.load() default to weights_only=True — fixes #40

Previously we were using sed to inject a compatibility patch into cloud_io.py.

Issues

• Fixes Workaround: SpeechBrain list_audio_backends() removed in torchaudio 2.9.x #38 — SpeechBrain list_audio_backends() removed in torchaudio 2.9.x
• Fixes Workaround: SpeechBrain torchaudio.info() removed in torchaudio 2.9.x #39 — SpeechBrain torchaudio.info() removed in torchaudio 2.9.x
• Fixes Workaround: Lightning weights_only default change in PyTorch 2.6+ #40 — Lightning weights_only default change in PyTorch 2.6+

Testing Required

This needs a full venv rebuild to verify:

rm -rf venv-nvidia
./scripts/install_venv.sh --nvidia
source setup_env.sh
source venv-nvidia/bin/activate
./scripts/process_single.sh <test-audio> --transcribers whisperx --processors opus

Verify that WhisperX transcription with speaker diarization works end-to-end without errors.

Not Changed

• Step 9 (pyannote.audio torchcodec fallback) — torchcodec v0.4 with CXX11 ABI fix has not been released yet (Workaround: torchcodec ABI incompatibility with PyTorch 2.9.1+cu130 #36, Document torchcodec ABI incompatibility workaround #44). Our soundfile fallback patch is still needed.