Neural TTS-Based Dynamic Data Augmentation for Improved Speech Separation

Text-to-speech (TTS) synthetic data augmentation has been widely used in various speech processing tasks, but its effectiveness in speech separation remains understudied. In this paper, we present SpeakerAugment+ (SA+), a neural TTS-based dynamic data augmentation framework for speech separation. The SA+ framework consists of two modules: speaker module, which learns a Gaussian Mixture Model (GMM) to characterize the distribution over speaker embeddings in training data and samples unseen speaker embeddings during inference; speech module, which conditions speech synthesis using speaker embeddings with controllable speaker parameters. SA+ incorporates three augmentation techniques: speaker generation, parameter manipulation and utterance generation, enhancing speaker and utterance diversity from different perspectives. Following the SA+ framework, we design FS2-SA+ and Matcha-SA+, which are based on FastSpeech 2 and Matcha-TTS, respectively. We evaluate SA+ across multiple separation models and datasets, and the results demonstrate a substantial improvement in speech separation performance. Matcha-SA+ generates higher-quality speech and achieves better separation performance in intra-corpus tests. Conversely, FS2-SA+ supports a broad range of speaker parameter adjustments, leading to better generalization. Besides, the relationship between harmonicity and speech separation is a widely researched topic, and our findings indicate that speech lacking an explicit harmonic structure, when generated by neural TTS, can function as augmented data to improve speech separation. This research underscores the effectiveness of neural TTS-based data augmentation in speech separation tasks. We hope that our work can offer insights for future studies investigating data augmentation strategies within speech separation.