Tampere University of Technology

TUTCRIS Research Portal

Deep neural network based speech separation optimizing an objective estimator of intelligibility for low latency applications

Research output: Chapter in Book/Report/Conference proceedingConference contributionScientificpeer-review


Original languageEnglish
Title of host publication16th International Workshop on Acoustic Signal Enhancement, IWAENC 2018
Number of pages5
ISBN (Electronic)9781538681510
Publication statusPublished - 2 Nov 2018
Publication typeA4 Article in a conference publication
EventInternational Workshop on Acoustic Signal Enhancement - Tokyo, Japan
Duration: 17 Sep 201820 Sep 2018


ConferenceInternational Workshop on Acoustic Signal Enhancement


Mean square error (MSE) has been the preferred choice as loss function in the current deep neural network (DNN) based speech separation techniques. In this paper, we propose a new cost function with the aim of optimizing the extended short time objective intelligibility (ESTOI) measure. We focus on applications where low algorithmic latency (≤ 10 ms) is important. We use long short-term memory networks (LSTM) and evaluate our proposed approach on four sets of two-speaker mixtures from extended Danish hearing in noise (HINT) dataset. We show that the proposed loss function can offer improved or at par objective intelligibility (in terms of ESTOI) compared to an MSE optimized baseline while resulting in lower objective separation performance (in terms of the source to distortion ratio (SDR)). We then proceed to propose an approach where the network is first initialized with weights optimized for MSE criterion and then trained with the proposed ESTOI loss criterion. This approach mitigates some of the losses in objective separation performance while preserving the gains in objective intelligibility.


  • Deep neural networks, Low latency, Speech intelligibility, Speech separation

Publication forum classification

Field of science, Statistics Finland