Robust speech recognition based on a Bayesian prediction approach

摘要: We study a category of robust speech recognition problem in which mismatches exist between training and testing conditions, no accurate knowledge the mismatch mechanism is available. The only available information test data along with set pretrained Gaussian mixture continuous density hidden Markov models (CDHMMs). investigate from viewpoint Bayesian prediction. A simple prior distribution, namely constrained uniform adopted to characterize uncertainty mean vectors CDHMMs. Two methods, model compensation technique based on predictive decision strategy called Viterbi classification are studied. proposed methods compared conventional decoding algorithm speaker-independent experiments isolated digits TI connected digit strings (TIDTGITS), where conditions caused by: (1) additive white noise, (2) each 25 types actual ambient noises, (3) gender difference. experimental results show that distribution techniques help improve performance robustness under examined conditions.