IVSE: Indian Voice Separation and Enhancement from a cocktail party scenario

Abstract

Audio bots like Alexa, Siri, Google assistant, require a clean voice to perform a task. These bots ignore a disturbance or mixed voice. We get the famous message “sorry I could not understand”. With the introduction of smart homes and smart cities, it is imperative for devices to understand the commands in a noisy environment and that too in a native language. Indian Voice Separation and Enhancement (IVSE) offer the solution. For separation and enhancement of voice, the model should filter the noise first. For eliminating zero and negative values, a Zero-Negative filter (ZNF) is created. To eliminate the rippling effect induced by the time domain or frequency domain filters, Enhance Voice Function (EVF) enhances the voice. The output is windowed into different frames of equal lengths. Then it is labelled as noise and clean signals. A gradient boosting algorithm-based approach is then applied to filter the noise. The 50,000 voiceprints from the filtered voice were used to construct a training and validation set. For predictive analysis, the dataset is divided into an 80:20 ratio. IVSE uses LightGBM to create distinctive voiceprints. LightGBM operates in the background on Tensor Flow, which improves its performance. The paper compares IVSE and different benchmark algorithms at the end.