The key to the new AI is that it’s unsupervised. Current machine learning systems ro applications such as speech processing need to be trained on data sets which associate human transcriptions to the raw speech files.
Because the system doesn’t require annotated data, it’s had to make a few assumptions. One is that words that occur in speech follow a standard power-law distribution. The team tested its system on six recordings of lectures given at MIT, but there were still problems. For instance, it thought that “open university” was one word, not two.
One big application for this kind of technology is that it might be able to be applied to languages that aren’t as widely spoken. It could also help the accuracy of systems that transcribe accented speech.