Improvements in connected digit recognition using linear discriminant analysis and mixture densities

Four methods were used to reduce the error rate of a continuous-density hidden Markov-model-based speech recognizer on the TI/NIST connected-digits recognition task. Energy thresholding sets a lower limit on the energy in each frequency channel to suppress spurious distortion accumulation caused by random noise. This led to an improvement in error rate by 15%. Spectrum normalization was used to compensate for across-speaker variations, resulting in an additional improvement by 20%. The acoustic resolution was increased up to 32 component densities per mixture. Each doubling of the number of component densities yielded a reduction in error rate by roughly 20%. Linear discriminant analysis was used for improved feature selection. A single class-independent transformation matrix was applied to a large input vector consisting of several adjacent frames, resulting in an improvement by 20% for high acoustic resolution. The final string error rate was 0.84%.<>