Model compensation using parallel model combination and adaptation technique for improved accuracy in noisy speech recognition

Abstract

Due to the mismatch between training and operating conditions, speech recognition systems often exhibit dramatic performance degradation when they are practically used in real-world environments. Various techniques based on noise resistance, speech enhancement and model compensation approaches have been widely used to improve the performance of speech recognition in noise. Parallel Model Combination (PMC) is one of the most popular techniques and has been proved powerful in compensating recognition models, so that they reflect speech in noisy acoustic environments. However, studies have shown that some assumptions and approximations made in the PMC, primarily in the domain transformation and parameter combination processes are not accurate in certain situations, thus restricting the achievement of better performance. This research suggests using Maximum Likelihood Spectral Transformation (MLST) as the adaptation technique to further improve the performance of PMC. MLST is a transformation-based technique. It has some advantages over other adaptation techniques like Maximum Likelihood Linear Regression (MLLR) and Maximum A Posterior (MAP). MLST requires only a small amount of adaptation data and the adaptation process is computationally inexpensive. The proposed method is denoted as Adaptive Parallel Model Combination (APMC).