上游的仅包含文本的推理报告

Summary

We attempted to run VibeVoice-1.5B for text-to-speech locally (without SGLang) using the published  vibevoice pip package. The model loads correctly and the tokenizer works after a minor patch, but the inference pipeline cannot produce audio without SGLang — the  generate() method's speech-processing path is never initialized for text-only input, and  speech_outputs returns  [None].