SPEAKER-DEPENDENT FEATURES FOR SPONTANEOUS SPEECH RECOGNITION

Abstract

This paper presents the results of the study on improving robustness to the acoustic variability of the speech signal for spontaneous speech recognition system. The method is proposed to constructing high-level bottleneck features using deep neural network adapted to a speaker and to acoustic environment with i-vectors. The proposed method provides 11,9% relative reduction of word error rate in Russian spontaneous telephone speech recognition task.

Keywords

• automatic speech recognition
• speaker adaptation
• i-vectors
• bottleneck features from deep neural network