Mixed Language Separation Using Deep Neural Network

With multiple languages spoken in the world by different groups of people, we may encounter mixed language speech to hear, especially while vlogging in a different country or during interviews with voice dubbing. The appropriate language speech audio can be extracted from a mixed one using a separation mechanism. This paper proposes a DNN model to perform such a language separation task. Different 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 as the input to the DNN. For the training target, the Short-Time Fourier Transform (STFT) Spectral Mask is considered. To understand the improvement on the speech, the processed speech is then evaluated for its intelligibility and quality. Here 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 the DNN. It can be observed from the results that the language separated audio using a trained DNN model has shown improved Intelligibility and Quality.