> My lab was the first, in 2001, to design such a filter, which labels sound streams as dominated by either speech or noise. With this filter, we would later develop a machine-learning program that separates speech from other sounds based on a few distinguishing features, such as amplitude (loudness), harmonic structure (the particular arrangement of tones), and onset (when a particular sound begins relative to others).
> Next, we trained the deep neural network to use these 85 attributes to distinguish speech from noise.
> One important refinement along the way was to build a second deep neural network that would be fed by the first one and fine-tune its results. While that first network had focused on labeling attributes within each individual time-frequency unit, the second network would examine the attributes of several units near a particular one
> Even people with normal hearing were able to better understand noisy sentences, which means our program could someday help far more people than we originally anticipated
> There are, of course, limits to the program’s abilities. For example, in our samples, the type of noise that obscured speech was still quite similar to the type of noise the program had been trained to classify. To function in real life, a program will need to quickly learn to filter out many types of noise, including types different from the ones it has already encountered
oh