Explore VibeVoice T2S model export & quantization

[IgorSwat]

Problem description

VibeVoice is an open-source Text to Speech model provided by Microsoft.

What should be done

1. Explore the model properties
2. Try to export the model to ExecuTorch format
3. Try to quantize the model

Benefits to React Native ExecuTorch

Compared to already implemented Kokoro model, it could bring some meaningful gains such as:

• Remove the need for using additional phonemization library
• Simplify the model's inference logic

[IgorSwat]

Update - model properties:

• Slow CPU inference, far too slow to consider a real-time mobile usage, at least not without a proper quantization
• Acceptable speed with MPS acceleration on Mac Book M4 Pro
• High memory consumption - the model itself weighs over 2GB - a quantization is pretty much essential.
• High quality speech output, possibly better than Kokoro
• Supports different languages, including Polish - without a need for additional phonemization

Update - model export:

• Unable to export the model with optimum-cli - the script seems to not recognize the model's architecture. I created an issue in the Optimum Executorch repository.
• Unable to download and load the model from HF repository through transformers package.

[msluszniak]

If the model export is currently impossible via Optimum ExecuTorch, I'll set up mentioned issue as a blocker for this one.