MTASS

音频源分离任务，例如语音增强、语音分离和音乐源分离，在最近的研究中取得了令人瞩目的表现。深度神经网络强大的建模能力让我们对更具挑战性的任务充满希望。本文发起了一项新的多任务音频源分离 (MTASS) 挑战，将语音、音乐和噪声信号从单声道混合中分离出来。首先，我们介绍了这项任务的细节，并生成了一个包含语音、音乐和背景噪声的混合数据集。然后，我们提出了一个复杂域中的 MTASS 模型，以充分利用三个音频信号的频谱特性差异。具体来说，所提出的模型遵循两级流水线，将三种音频信号分离，然后分别进行信号补偿。在比较不同的训练目标后，选择复比掩模作为MTASS更合适的目标。实验结果还表明，残差信号补偿模块有助于进一步恢复信号。与几种众所周知的分离模型相比，所提出的模型在分离性能方面显示出显着优势。Audio source separation tasks, such as speech enhancement, speech separation, and music source separation, have achieved remarkable performance in recent research. The powerful modeling capabilities of deep neural networks have instilled optimism for tackling more challenging tasks. In this paper, we launch a novel multi-task audio source separation (MTASS) challenge that focuses on separating speech, music, and noise signals from monaural mixtures. First, we introduce the details of this task and generate a mixed dataset containing speech, music, and background noise. Subsequently, we propose an MTASS model in the complex domain to fully leverage the differences in spectral characteristics among the three audio signals. Specifically, the proposed model follows a two-stage pipeline: it first separates the three audio signals, then performs signal compensation individually. After comparing different training objectives, complex ratio masking is selected as the more suitable target for MTASS. Experimental results also demonstrate that the residual signal compensation module helps further recover the audio signals. Compared with several well-known separation models, the proposed model exhibits significant advantages in separation performance.

搜集汇总

背景与挑战

背景概述

MTASS是一个用于多任务音频源分离的数据集，包含语音、音乐和背景噪声的混合音频，旨在从单声道混合中分离出这三种信号。该数据集支持相关研究，并提出了一个复杂域的两阶段模型，在分离性能上表现出显著优势。