VibeVoice ExLlamaV3 fork speeds up local ASR

Strong practical win if you want local speech models with less latency and better throughput, especially on hardware where quantized inference matters more than absolute fidelity.

• –The main value here is speed: q8 quantization plus ExLlamaV3 appears to unlock a meaningful runtime improvement over the standard Transformers fp16 path.
• –This is most relevant for people already using VibeVoice locally and willing to trade some simplicity and possibly some quality for performance.
• –It reads like an experimental fork, not an official upstream release, so stability, compatibility, and maintenance are the main unknowns.
• –The post is light on benchmark methodology, so the “4x faster” claim should be treated as indicative rather than definitive.