Abstract
Recently, research on Universal Sound Separation (USS) has shifted focus to Language-queried Audio Source Separation (LASS), which is more aligned with human usage patterns. A key challenge faced by LASS when separating arbitrary sounds is the substantial variations in feature distributions across distinct sound classes, which renders it challenging to effectively model all audio classes. Inspired by the pre-training + fine-tuning paradigm in large language models (LLMs), we propose a novel LASS method that integrates domain partitioning and parameter-efficient fine-tuning (PEFT). Our method consists of three stages: 1) Domain partitioning: We use the K-Means algorithm to cluster the mean embeddings of each sound class within the training set, resulting in multiple data subdomains. 2) Fine-tuning: We apply PEFT to the pre-trained AudioSep model, fine-tuning it on each subdomain separately, where each subdomain corresponds to a dedicated PEFT module based on Conv-Adapter. 3) Inference: A subdomain classifier is designed to route the test mixture and text prompt to the correct PEFT module. Experimental results on six evaluation benchmarks show that our method achieves an average SDRi of 9.76 dB and SI-SDR of 9.06 dB, outperforming the baseline AudioSep by 1.01 dB and 1.29 dB, respectively. Our code has been open-sourced1.