Building noise suppression module for voice signal on OMICRM platform at VIHAT

Abstract

In the digital age, excellent customer service is critical for increasing customer retention, which is one of the most effective ways to ensure a business' profitability. Customers always expect more, and providing customer support as a service is critical to meeting those expectations. While phone-based customer service remains popular, call centers are rapidly growing in popularity. When designing a system in the context of audio communication for transmission purposes, noise influences must be carefully considered. In this thesis, we apply a module for improving real-time speech enhancement of virtual phone calls using an LSTM-core model (Long Short-Term Memory [1]), an artificial recurrent neural network used in the fields of artificial intelligence and deep learning. LSTMs are designed to revoke the long-term dependency issue, remembering past information for a long period of time is generally their default way of behaving. The goal of training is to produce an algorithm that reduces noise and enhances human voice in a noisy audio signal. The material used for analysis and comparison is both in English and Vietnamese, but only the Vietnamese data is used to train the final outcome model. The model was able to suppress some common ambient noises in our daily surroundings.