In the mid-1980s, Terrence Sejnowski and Charles Rosenberg conducted research that resulted in the creation of NETtalk, an artificial neural network. The objective of this program was to develop simplified models that could shed light on the intricacy of human-level cognitive tasks and implement a connectionist model that could learn to perform similar tasks.
NETtalk was trained to pronounce written English text by comparing phonetic transcriptions with the input text. The program was trained on a vast amount of English words and their corresponding pronunciations, enabling it to accurately generate pronunciations for unseen words. Its success inspired further research in speech synthesis and pronunciation generation and highlighted the potential of neural networks for solving complex NLP problems.
The network was designed to handle the complexities of English, including its irregular spelling-to-sound relationships, and was trained in an unsupervised manner, without the use of any annotated data.