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

TL;DR

This paper introduces a novel method for creating a realistic speech separation dataset with ground truths by recording two speakers simultaneously, enabling better training and benchmarking of deep learning models in real-world scenarios.

Contribution

The paper presents a new approach to generate realistic speech separation datasets with ground truths, addressing the challenge of obtaining accurate source signals in natural environments.

Findings

Improved SI-SDR by 1.65 dB using the new dataset

Enhanced PESQ score by approximately 0.5

Method increased model stability at various microphone-speaker distances

Abstract

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.