A COMPARISON OF PREPROCESSORS FOR THE CAMBRIDGE RECURRENT ERROR PROPAGATION NETWORK SPEECH RECOGNITION SYSTEM
Tony Robinson, John Holdsworth, Roy Patterson and Frank Fallside
September 1990
This paper makes a comparison of several preprocessors for the task of speaker independent phoneme recognition from the TIMIT database using a recurrent error propagation network recogniser (Robinson Fallside 90, TR42)
The paper evaluates FFT, filterbank, auditory model and LPC based techniques in the spectral and cepstral domains and adds some simple features such as estimates of the degree of voicing, formant positions and amplitudes. The paper concludes that the features do not make a significant contribution and that the spectral domain representations, independent of their derivation, are better suited to this task. However, we find that the recogniser was relatively insensitive to preprocessor and changes in the architecture and training of the recogniser are more significant.
The current recognition rate on the TIMIT database of 61 symbols is 69.5% correct (64.0% including insertion errors) and on a reduced 39 symbol set the recognition rate is 76.1% correct (70.4%). This compares favourably with the results of other methods, such as Hidden Markov Models, on the same task.
If you have difficulty viewing files that end '.gz',
which are gzip compressed, then you may be able to find
tools to uncompress them at the gzip
web site.
If you have difficulty viewing files that are in PostScript, (ending
'.ps' or '.ps.gz'), then you may be able to
find tools to view them at
the gsview
web site.
We have attempted to provide automatically generated PDF copies of documents for which only PostScript versions have previously been available. These are clearly marked in the database - due to the nature of the automatic conversion process, they are likely to be badly aliased when viewed at default resolution on screen by acroread.