Clean Speech: To separate speech from noise, a machine learning program breaks a noisy speech sample into a collection of elements called time-frequency units. Next, it analyzes these units to extract 85 features known to distinguish speech from other sounds. Then, the program feeds the features into a deep neural network trained to classify the units as speech or not based on past experience with similar samples. Lastly, the program applies a digital filter that tosses out all the nonspeech units to leave only separated speech.