Enhancement of Assamese Speech Signals Using Learning Based Techniques

For efficient recognition and proper interpretation of a speech signal in ASR models, de-noising of corrupted signals becomes very important. There are many traditional noise removing techniques which proved to be reliable but researchers have found the learning based approaches to be more fruitful than those traditional methods. Deep learning techniques have been popularly used in speech de-noising and recognition purposes and have become an integral part of such ASR systems. In this work, Auto-encoders and other ANN models have been implemented to remove noise from Assamese speech signals irrespective of the speaker, gender and the dialectal identity of the samples. The sample set consist of clean Assamese sentences of two, three, four, five and six words and corrupted signals with AWGN variation of 3 dB to -3 dB which increases the sample size and also makes the system robust. Despite of certain limitations, the satisfactory experimental results justifies the proposed de-noising method for Assamese speech.