Hopfield Model Applied to Vowel and Consonant Discrimination

摘要

In the preliminary experiments described in this report, Hopfield networks have proved to be fairly robust implementations of associative memory modules. For speech-processing applications, an appropriate binary representation of the speech is central for obtaining good performance. For different applications (e.g., recognition, vocoding, pitch detection) the representation is bound to be different. Our attempt to devise an automatic DRT with Hopfield networks indicates that incorporating null states could significantly enhance performance. In a rhyme test such as DRT, the critical interval of the speech is at or near the consonant-vowel boundary. If the noncritical intervals could be made to converge to such a null state while the critical regions converged to a true memory state, much of the human editing could be eliminated. Several proposals have been made (but not yet tested) on this subject: (a) introduce new phony states orthogonal to the real states and to each other; convergence to any phony state would be interpreted as a null state; (b) introduce a bias in the threshold computation. The experiments described in this report are only a beginning toward finding useful applications of neural-network-like structure. Experimental and theoretical studies of such networks have high priority in future efforts in this new and interesting area. Keywords: Automatic Diagnostic Rhyme Test.