Developing an Effective Training Dataset to Enhance the Performance of AI-based Speaker Separation Systems

This paper addresses the challenge of speaker separation, which remains an active research topic despite the promising results achieved in recent years. These results, however, often degrade in real recording conditions due to the presence of noise, echo, and other interferences. This is because neural models are typically trained on synthetic datasets consisting of mixed audio signals and their corresponding ground truths, which are generated using computer software and do not fully represent the complexities of real-world recording scenarios. The lack of realistic training sets for speaker separation remains a major hurdle, as obtaining individual sounds from mixed audio signals is a nontrivial task. To address this issue, we propose a novel method for constructing a realistic training set that includes mixture signals and corresponding ground truths for each speaker. We evaluate this dataset on a deep learning model and compare it to a synthetic dataset. We got a 1.65 dB improvement in Scale Invariant Signal to Distortion Ratio (SI-SDR) for speaker separation accuracy in realistic mixing. Our findings highlight the potential of realistic training sets for enhancing the performance of speaker separation models in real-world scenarios.

Towards Solving Cocktail-Party: The First Method to Build a Realistic Dataset with Ground Truths for Speech Separation

Speech separation is very important in real-world applications such as human-machine interaction, hearing aids devices, and automatic meeting transcription. In recent years, a significant improvement occurred towards the solution based on deep learning. In fact, much attention has been drawn to supervised learning methods using synthetic mixtures datasets despite their being not representative of real-world mixtures. The difficulty in building a realistic dataset led researchers to use unsupervised learning methods, because of their ability to handle realistic mixtures directly. The results of unsupervised learning methods are still unconvincing. In this paper, a method is introduced to create a realistic dataset with ground truth sources for speech separation. The main challenge in designing a realistic dataset is the unavailability of ground truths for speakers signals. To address this, we propose a method for simultaneously recording two speakers and obtaining the ground truth for each. We present a methodology for benchmarking our realistic dataset using a deep learning model based on Bidirectional Gated Recurrent Units (BGRU) and clustering algorithm. The experiments show that our proposed dataset improved SI-SDR (Scale Invariant Signal to Distortion Ratio) by 1.65 dB and PESQ (Perceptual Evaluation of Speech Quality) by approximately 0.5. We also evaluated the effectiveness of our method at different distances between the microphone and the speakers and found that it improved the stability of the learned model.