This document summarizes a research paper on determining noise levels in noisy speech signals. It discusses extracting amplitude modulation spectrogram (AMS) features from noisy speech segments and classifying them as noise-dominated or signal-dominated using Bayes' classifier. The classified features are reconstructed using overlap-and-add to estimate the signal-to-noise ratio between the original clean speech and reconstructed speech, and between the original clean speech and noisy speech. The paper presents results on speech segments corrupted with different noises, showing SNR is reduced between the original clean and reconstructed speech compared to the original clean and noisy speech.