Next: The Recurrent Network
Up: THE BASIC SYSTEM
Previous: The TIMIT Database
For practical reasons, (memory limitations on disk and in the transputer RAM), the preprocessed data was scaled to fit into 8 bits per channel. This was done by computing a histogram and scaling so that no more than one in 500 samples lies outside the central 15/16th of the range. Typically this meant that one sample in 1000 would be thresholded.