Reproduce “On The Landscape of Spoken Language Models: A Comprehensive Survey” 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
Siddhant Arora, Kai-Wei Chang, Chung-Ming Chien, Yifan Peng, Haibin Wu, Yossi Adi, Emmanuel Dupoux, Hung-Yi Lee and 2 more
This survey paper maps the evolving landscape of Spoken Language Models (SLMs) by unifying terminology, categorizing diverse architectures and training methodologies, and detailing current evaluation practices. It highlights the field's progression toward universal speech processing systems and identifies persistent challenges and future research directions.
Key insights
- SLMs are increasingly adopting multi-stage training paradigms, including generative pre-training and instruction tuning, to enhance task versatility, often leveraging synthetic data.
- The choice and integration of speech representations, from continuous features to discrete phonetic or audio codec tokens, critically influence an SLM's capacity for both linguistic understanding and high-fidelity speech generation.
- Natural, duplex conversational interaction requires specialized architectural solutions (e.g., dual-channel, time-multiplexing) to handle simultaneous listening and speaking, turn-taking, and interruptions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.