NETtalk (artificial neural network)

NETtalk is an artificial neural network that was created mid-80s by Terrence J. Sejnowski and Charles Rosenberg and written ( English ) text ( ie be converted by speech synthesis graphemes into phonemes it ) into a coding of the debate.

Construction

NETtalk is constructed of three layers Multilagenperzeptron with seven groups of 29 neurons in the input, 80 neurons in the hidden and 26 neurons in the output layer. Each of the groups in the input layer encodes a letter of the input word ( the 29 neurons correspond to the 26 letters of the alphabet and each one neuron for space, end of sentence and other punctuation ), the fourth group represents this letter, its associated phoneme is to determine the network and the remaining groups are the essential for the correct determination of the context of the three previous or subsequent letters dar.

For training the network correct grapheme -phoneme combinations were used, so it is a method for supervised learning.

Performance

After 50 training runs on a data set of 1024 words, the network reached an accuracy of 95 % on the training and 78 % on the test data.

Influence

In the 1980s represented a NETtalk the sensational applications that prompted many scientists again to carry out research in the field of connectionism. Critics doubt, however, that this ( be similar successes with ' conventional ' programs achieved ) on the quality of the architecture was. Rather, reference is made to the presentation of the learning process of the network: The output from the network phonemes were issued as a spoken language, the program began with incomprehensible juxtaposition of sounds and improved gradually to understand language. In addition, for this presentation, a voice was used with a high pitch, so that there was for the audience the impression that a child learning to speak.

Sound

Http://www.cnl.salk.edu/Media/nettalk.mp3