Developed for Finnish (a phonetic language, written as it is said)
Trained on one speaker, will generalise to others.
A neural network is trained to cluster together similar sounds, which are then labelled with the corresponding character.
When recognising speech, the sounds uttered are allocated to the closest corresponding output, and the character for that output is printed.
requires large dictionary of minor variations to correct general mechanism
noticeably poorer performance on speakers it has not been trained on