Reproduce “GLM-4-Voice: Towards Intelligent and Human-Like End-to-End Spoken Chatbot” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Aohan Zeng, Zhengxiao Du, Mingdao Liu, Kedong Wang, Shengmin Jiang, Lei Zhao, Yuxiao Dong, Jie Tang
GLM-4-Voice is an end-to-end spoken chatbot from Zhipu.AI and Tsinghua University that unifies speech and text processing within an auto-regressive transformer. The model leverages an ultra-low bitrate speech tokenizer and massive speech-text pre-training to enable intelligent, human-like voice conversations with reduced latency.
Key insights
- Bridging text and speech modalities through large-scale pre-training effectively enables LLM intelligence in spoken chatbots.
- A single-codebook, ultra-low bitrate speech tokenizer (175 bps) facilitates efficient autoregressive generation and knowledge transfer from text-based LLMs.
- A two-stage training process, combining massive speech-text pre-training with specialized supervised fine-tuning using a 'streaming thoughts' template, yields intelligent and low-latency spoken interactions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.