In the early days of automatic speech recognition, systems were so prone to error that even using a different microphone or acoustic channel for training and testing speech recognition systems could lead to failure. Clearly, systems would need to tolerate a lot more variability if they were to enable practical applications.
To overcome this hurdle, ICSI’s Nelson Morgan and Hynek Hermansky, then of the Oregon Graduate Institute, partnered to create a simple filtering approach that removed the effect of a constant linear spectral characteristic in the audio. Their solution, called RASTA, reduced the error rate for a digit recognition task from 60% to a mere 3%. Although their technique wasn’t universally adopted, it raised the profile of the problem and within a few years, every research system for speech recognition either used RASTA or an alternate fix for the issue—helping to move the field forward toward the ubiquitous speech recognition applications we enjoy today.
What made ICSI a good place to pursue these projects?
“There are three things that come to mind. First, ICSI encouraged “curiosity-driven research.” Researchers pursued whatever they wondered about, which led to many serendipitous discoveries. Secondly, as part of the broader Berkeley intellectual environment, at ICSI there were incredible opportunities to interact with other curious minds, often reaching across multiple fields; this included many wonderful visitors, many from other countries, whose differing backgrounds and interests further “stirred the pot.” And finally, since ICSI itself had no teaching requirement, unless someone chose to connect with campus in this way, one’s time was totally devoted to research.”
Nelson Morgan
Director, Usable Security and Privacy
This story was published in January 2026 as part of a retrospective series highlighting ICSI’s accomplishments and impacts over the years. To learn about our ongoing work, explore our Core Research Themes.