Large-Scale Learning on Overlapped Speech Detection: New Benchmark and New General System

TL;DR

This paper introduces a large-scale learning approach for overlapped speech detection, proposing a new benchmark and a general system that significantly improves accuracy and robustness in diverse acoustic environments.

Contribution

It presents a new large-scale benchmark dataset and a general OSD system based on Conformer networks with large-scale learning, advancing robustness and accuracy.

Findings

LSL significantly improves OSD accuracy and robustness.

CF-OSD outperforms existing systems on the new benchmark.

Achieves state-of-the-art results on small dataset benchmarks.

Abstract

Overlapped Speech Detection (OSD) is an important part of speech applications involving analysis of multi-party conversations. However, most of existing OSD systems are trained and evaluated on small datasets with limited application domains, which led to the robustness of them lacks benchmark for evaluation and the accuracy of them remains inadequate in realistic acoustic environments. To solve these problem, we conduct a study of large-scale learning (LSL) in OSD tasks and propose a new general OSD system named CF-OSD with LSL based on Conformer network and LSL. In our study, a large-scale test set consisting of 151h labeled speech of different styles, languages and sound-source distances is produced and used as a new benchmark for evaluating the generality of OSD systems. Rigorous comparative experiments are designed and used to evaluate the effectiveness of LSL in OSD tasks and define the OSD model of our general OSD system. The experiment results show that LSL can significantly improve the accuracy and robustness of OSD systems, and the CF-OSD with LSL system significantly outperforms other OSD systems on our proposed benchmark. Moreover, our system has also achieved state-of-the-art performance on existing small dataset benchmarks, reaching 81.6\% and 53.8\% in the Alimeeting testset and DIHARD II evaluation set, respectively.