LGAS: Language-Guided Audio Synthesis via Dual-Path Diffusion and Semantic Modeling

Recent advancements in music generation have been marked by significant progress, particularly with the introduction of MusicLM, a state-of-the-art system employing a threetiered language model to sequentially process semantic, coarse acoustic, and fine acoustic representations.Despite its impressive performance, MusicLM's reliance on multi-stage processing renders it computationally expensive and unsuitable for real-time applications.To address this, we propose LGAS (Language-Guided Audio Synthesis), a novel framework that significantly reduces generation time by up to 95.7% for 10-second audio and 99.6% for 30second audio while maintaining comparable output quality.LGAS retains the highest-level semantic modeling of MusicLM and introduces a dual-path diffusion (DPD) architecture integrated with an audio VAE-GAN decoder.This design enables simultaneous capture of both coarse and fine acoustic features via cross-attention at each denoising step, effectively embedding semantic information into the latent space.Experimental results show that LGAS outperforms prior models in generation speed, continuity, musical fidelity, and prompt relevance, setting a new benchmark for efficient text-to-music generation.