Create speaker embeddings fix

[michalkulakowski]

🛠 Summary

JIRA/Issue if applicable.
Describe the changes.

🧪 Checklist

• Unit tests added.
• The documentation updated.
• Change follows security best practices.
  ``

[Copilot]

Pull request overview

Updates the Python dependency set for the demos/audio flow (notably the speaker-embedding preparation step) by adjusting pinned versions and adding missing direct dependencies required during runtime/model download.

Changes:

• Downgrade pinned torch / torchaudio versions used by the audio demo requirements.
• Adjust speechbrain version and add huggingface_hub / requests requirements.

[Copilot]

This file pins torch/torchaudio to 2.5.1+cpu, but the audio demo README later installs demos/common/export_models/requirements.txt in the same environment, which pins torch==2.9.1+cpu (and does not pin torchaudio). Installing both requirement sets sequentially will force pip to upgrade/downgrade torch and can break either SpeechBrain or the export script depending on the final resolved versions. Consider aligning the torch (and torchaudio) versions across both requirements files, loosening the pin to a compatible range, or updating the demo instructions to use separate virtual environments for the two steps.

Suggested change

	torch==2.5.1+cpu
	torchaudio==2.5.1+cpu
	torch==2.9.1+cpu
	torchaudio==2.9.1+cpu

[Copilot]

huggingface_hub<1.0 leaves the exact installed version floating, which can make the demo non-reproducible over time and can change behavior of model downloads unexpectedly. If a specific version is known to work with speechbrain==1.0.2, consider pinning it (or at least adding a tested minimum bound) to stabilize installs.

Suggested change

	huggingface_hub<1.0
	huggingface_hub==0.29.3