From NICK RYMAN-TUBB
Your wonderful front page graphics and title promised us ‘Computers
that listen’ (4 December), but the article seems to have been deaf to the
amazing work done using neural computers.
Neural computers were inspired by research into how the biological brain
works. In a discontinuous speech recognition system a neural network approach
performs as well as or better than hidden Markov modelling (HMM). However,
it outperforms HMM in all cases with continuous speech recognition. It does
not require an endpointing algorithm and can model the coarticulation between
words. The neural network makes use of context information that is unavailable
to frame-level HMM processing. What is more important, a neural network
approach does not require the ‘ten times computing power’ mentioned in
your article.
First there were bandpass filters in the 1960s, linear predictive coding
in the 1970s and then HMM in the 1980s. Managers have invested company resources
in these technologies through the decades to crack the speech recognition
problem. How can we now persuade these companies to throw away this work
and move to neural computing – the technology of the 1990s?
Nick Ryman-Tubb Petersfield, Hampshire
Advertisement
