Adaptive Parallel Model Combination for reduced environmental mismatch in noisy speech recognition

Due to environmental mismatch, speech recognition systems often exhibit drastic performance degradation in noisy conditions. This paper presents a model based technique termed adaptive parallel model combination (APMC) which compensates the initial acoustic models to reduce the discrepancy. APMC used the well-known PMC technique to composite a set of corrupted speech models, while fine tuning the mean parameter of the models using a transformation-based adaptation technique called Maximum Likelihood Spectral Transformation (MLST). Evaluated on a context-independent phone recognition task, APMC was found to be superior to both PMC and MLST, especially in non-stationary noisy conditions. On average, APMC has achieved 48.81% improvement over the initial models, whereas PMC and MLST have improved the accuracy by 34.12% and 35.23% respectively.