Bi-LSTM Model for Speech Recognition

For voice synthesis to sound realistic and understandable, fundamental frequency (F0) modeling is essential. Here, we work to improve F0 modeling for Isarn's voice synthesis system. For this problem, we introduce the F0 model, a Recurrent Neural Network (RNN) that has Bi-Long short-term memory (Bi-LSTM). Together with their dynamic features, the F0 values gathered from continuous Isarn speech at the statement level are used to highlight the supra-segmental aspects of the F0 contour. To get optimal performance, we investigate several deep RNN architectures and feature combinations for language processing. We used comparisons to numerous RNN-based baselines to evaluate the suggested technique. Both the state-of-the-art RNN approach for predicting F0 values at the picture level and the baseline RNN model for visualizing the F0 contour of syllables through a discrete cosine transform were surpassed by the suggested model.