Mixed Language Separation Using DNN

Abstract

With multiple language spoken in the world by the different group of people, it is possible that we may encounter with such mixed language speech to hear, especially during data share in an audio-video format. Such blended audio is not clearly understandable to the audience. Hence it is needed to be processed with some separation method. Mixed Language separation is the task of separating a particular language speech from a mixed language speech audio. Deep neural network (DNN) based speech enhancement algorithm shows good results in the non-stationary environment and can successfully outperform conventional methods which are based on unsupervised algorithms. All the information in the acoustic signal is too cumbersome to deal with, hence speech signal is illustrated by predetermined number of components of the signal which named as feature of the speech signal. Features like Mel Frequency Cepstrum Coefficient (MFCC), Power Spectrum and Relative Spectral Transformed Perceptual Linear Prediction coefficient (RASTA-PLP) are extracted from the mixed language speech, which then used as the input to the DNN. For training target, the Short-Time Fourier Transform (STFT) Spectral Mask is considered. Short-Time Objective Intelligibility (STOI) and Perceptual Evaluation of Speech Quality (PESQ) scores are used to compare the Intelligibility and Quality of the separated language speech signal processed by DNN. Language separated speech using trained DNN model has been found to achieve improve Intelligibility and Quality.