NETtalk learns to read aloud

NETtalk was a three-layer network with 309 units and about 18,000 weights, trained by backpropagation to map a window of seven letters to the phoneme of the middle one. Terrence Sejnowski and Charles Rosenberg trained it on a thousand words of transcribed speech and then fed its output through a speech synthesiser. The recording that resulted was the demonstration the connectionist revival needed. At first the machine produces a stream of babble; after a night of training it reads text aloud in a voice that is childlike and largely correct.

The comparison people drew was to DECtalk, the commercial rule-based synthesiser that had taken linguists years to build by hand. NETtalk had reached most of the way there from nothing, in hours, by example. The point was not the speech, which was worse than DECtalk's, but the method: knowledge that experts had spent careers writing down could be learned.

The network also let Sejnowski look inside. Hidden units clustered vowels apart from consonants without being told what those were, an early demonstration that learned representations could be inspected and would sometimes correspond to human categories. Interpretability research in the 2020s does the same thing on networks a million times larger.