VibeVoice-ASR-BitNet Technical Report
Published in arXiv preprint arXiv:2607.21075, 2026
VibeVoice-ASR-BitNet is a compressed variant of VibeVoice-ASR optimized for real-time inference on edge CPUs. It applies full-pipeline INT8 quantization to the VAE acoustic tokenizer and BitNet-style ternary weights to the autoregressive language model, while progressive quantization-aware training preserves accuracy. Custom SIMD kernels and fused ggml operators target both ARM and x86 platforms, achieving real-time recognition on low-thread-count CPUs.
Recommended citation: S. Xu, T. Song, S. Huang, Z. Peng, Y. Xia, Y. Tu, X. Huang, X. Wu, W. Wang, Y. Chang, J. Yu, L. Dong, and F. Wei. (2026). "VibeVoice-ASR-BitNet Technical Report." arXiv preprint arXiv:2607.21075.
Download Paper