“Asr-project-segment”版本间的差异

2017年11月16日 (四) 05:03的版本

Introduction

Speaker segmentation is important for many applications, among which include speaker-dependent adaptation, telephone archive analysis. Traditional approaches include ergodic HMM re-estimation, turn point detection and clustering, i-vector clustering. All these methods, however, are highly vulnerable for noise corruptions, speech overlapping, data imbalance.

We developed a deep segmentation approach that is based on deep learning approach that can analysis the true underlying speaker properties of speech signals, and then use simple clustering methods to achieve very high accuracy in segmentation.

Demonstration

A demo can be found <http://47.92.96.222/display/demo/?button=Call1_18_44741326_1_26 here>

<img src=http://cslt.riit.tsinghua.edu.cn/mediawiki/images/2/22/Seg.png>