Paper Title
LSTM based Kalman Filter for Punjabi Speech Enhancement

Abstract
We used a deep learning model called Long Short Term Memory to implement Speech Enhancement For Punjabi (LSTM). We carry out our work by training two LSTMs separately. The goal of training two LSTMs is to map learning by both LSTMs, with one learning from noisy acoustic features to clean speech signal magnitude and the second learning from noisy acoustic features to Line Spectrum Frequencies (LSFs). And after the mapping, the noisy speech phase is fused with the estimated clean magnitude for the reconstruction of the estimated clean speech. Following that, LSFs are converted into LPCs. Punjabi reconstructed speech, which is KF and aids in noise elimination. Acoustic and tonal features are extracted together with acoustic features. The evaluation parameter was the Word Error Rate (WER). In comparison to other combinations used in our work, MFCC+GFCC+BFCC+pitch provided the best WER of 21.33%. Keywords - Speech ehnacement, Punjabi ASR, Deep learning , Kalman Filter, Punjabi Tones, , Long Short Term Memory.