NETtalk Neural Network Developed by Terrence J. Sejnowski and Charles Rosenberg
Terrence J. Sejnowski of the Salk Institute and Charles Rosenberg of Princeton University developed NETtalk, a feedforward neural network trained to convert English text to speech, publishing the principal account in Complex Systems in 1987. The network learned pronunciation from examples alone, demonstrating that a multi-layer perceptron could acquire a complex linguistic skill without hand-coded rules.
Background
Through the 1970s and into the 1980s, getting a computer to read English text aloud meant writing rules by hand. Linguists and engineers would sit down together and try to encode the logic of English pronunciation: this letter combination makes this sound, except when it follows that pattern, except when the word is borrowed from French. The rule sets grew large and brittle. English spelling is irregular enough that every new exception required another patch, and the systems still stumbled on words they had never seen.
Neural networks were a potential alternative, but they had been largely set aside after early enthusiasm faded. The perceptron, a simple single-layer network, had been shown in 1969 by Marvin Minsky and Seymour Papert to be unable to learn certain basic functions. Multi-layer networks could in principle do more, but nobody had a reliable way to train them. The backpropagation algorithm, which adjusts a network’s internal weights by working backwards from its errors, had been developed and refined through the early 1980s, but it had not yet been applied to a problem complex enough to make people sit up.
Pronunciation was exactly that kind of problem. It is not random, but it is not fully systematic either. A child learns it from hearing thousands of examples, not from studying a rulebook.
What happened
Terrence J. Sejnowski, then at the Salk Institute for Biological Studies, and Charles Rosenberg, then a graduate student at Princeton University, built NETtalk to see whether a network trained purely on examples could pick up that skill. The network they described in Complex Systems in 1987 was a feedforward multi-layer perceptron: information flowed in one direction, from input to output, through one or more hidden layers of units sitting between the two. There were no cycles, no memory of previous states. The network simply mapped inputs to outputs and learned from its mistakes.
The input was a window of seven consecutive letters from a piece of text, centred on the letter whose sound the network was trying to predict. The output was a phoneme, the unit of sound in spoken language, along with stress information. During training, the network used backpropagation to adjust its weights each time it got the pronunciation wrong, guided by a corpus of continuous speech from a child. Sejnowski and Rosenberg used a text of about 1,000 words for training and tested on separate material.
What made NETtalk memorable was partly audible. As training progressed, the output was fed to a speech synthesiser in real time. At the start, the network produced babble. After a few passes through the training data, something that sounded like speech emerged, though the words were not yet right. By the end, it was reading the text with reasonable accuracy. You could hear the network learning, which was unusual and striking in a way that a table of numbers was not. The published account in Complex Systems described the network’s performance and its internal representations, including the finding that the hidden units developed structured groupings that loosely corresponded to phonetic categories, a result that drew interest from people studying both artificial networks and human cognition.
The work came out of Johns Hopkins University as a technical report in 1986 before the journal paper appeared in 1987. Both Sejnowski and Rosenberg had connections to Johns Hopkins alongside their Salk and Princeton affiliations at the time.
Why it mattered
NETtalk provided one of the most visible early demonstrations that backpropagation could train a multi-layer network to perform a cognitively demanding, rule-governed task, converting written English to phonemes, purely from data. Because its learning progress was audible, it became a widely cited illustration of connectionist learning for both specialist and general audiences, helping to rehabilitate neural network research after the so-called AI winter. It also stimulated debate about the degree to which distributed representations in artificial networks could model aspects of human skill acquisition.
People
Terrence J. Sejnowski, Chuck Rosenberg
Organisations
Salk Institute for Biological Studies, Princeton University, Johns Hopkins University
Sources
- Parallel Networks that Learn to Pronounce English Text.Complex Systems Publications.Primary source
- Parallel Networks that Learn to Pronounce English Text (Salk Institute author copy).Computational Neurobiology Laboratory, Salk Institute.Official
- NETtalk technical report (Johns Hopkins University EECS TR-86/7).Johns Hopkins University / CiteSeerX.Secondary
Cite this page
AI Achievements. (1987). NETtalk Neural Network Developed by Terrence J. Sejnowski and Charles Rosenberg. Retrieved 2026-08-22, from https://achievements.ai/milestone/nettalk-neural-network-by-terrence
@misc{achievements_nettalk_neural_network_by_terrence,
title = {NETtalk Neural Network Developed by Terrence J. Sejnowski and Charles Rosenberg},
author = {{AI Achievements}},
year = {1987},
url = {https://achievements.ai/milestone/nettalk-neural-network-by-terrence}
}