This paper introduces UAF (Unified Audio Front-end LLM), a novel large language model that unifies critical audio front-end tasks like voice activity detection, speaker recognition, and automatic speech recognition into a single end-to-end generative framework. UAF aims to overcome the limitations of traditional cascaded pipelines and enhance full-duplex speech interaction by jointly modeling semantic content and interaction-level control signals.

SLAM-LLM is an open-source deep learning framework designed to train customized Multimodal Large Language Models (MLLMs), with a focus on speech, language, audio, and music processing. It provides a modular configuration, detailed training and inference recipes, and high-performance checkpoints for mainstream tasks, aiming to accelerate research in audio-language models.