کنفرانس مهندسی برق ایران

صفحه اصلی / سی و سومین کنفرانس بین المللی مهندسی برق

Single-Channel Recursive Speech Separation with Unknown Speaker Count by Mask Estimation

نویسندگان :

Hadi Alizadeh¹ Rahil Mahdian Toroghi² Hassan Zareian³

1- Iran Broadcasting University 2- Iran Broadcasting University 3- Iran Broadcasting University

کلمات کلیدی :

Instantaneous Speech separation،single microphone،unknown speaker count،recursive operation،Mask estimation

چکیده :

This paper presents a novel speech separation method capable of handling an unknown number of speakers using a single, compact model, eliminating the need for prior knowledge of speaker count. The proposed approach employs a unique objective function to train a speaker-independent, single-channel model, enabling effective separation across diverse conditions, even when training and testing datasets differ. Additionally, a robust technique for detecting the number of speakers in a mixture is introduced, ensuring high performance with minimal computational complexity. By employing a recursive separation strategy, the method addresses the limitations of traditional approaches reliant on predefined speaker counts, making it more adaptable to real-world scenarios. Evaluations on the WSJ0 dataset demonstrate the proposed model's superiority in SI-SNR and SDR metrics while achieving a significantly lower parameter count compared to existing methods.