Evaluation of HMM-Based Feature Compensation Applied to a Finite-State Grammar Speech Recognizer

In this paper, we describe a Hidden Markov Model (HMM)-based feature-compensation method. The proposed method compensates for noise-corrupted features using the output prob-ability density functions (pdfs) of clean acoustic HMMs pro-vided to the recognizer in advance. In this way, the proposed method achieves model-based feature compensation without any extra parameters. In compensating for the features, the output pdfs are adaptively weighted according to forward path probabilities. Because of this, the proposed method can mini-mize degradation of feature compensation accuracy due to tem-porary changes in the noise environment. We applied the pro-posed feature compensation to a finite-state grammar speech recognizer and evaluated it by conducting hundred-word recog-nition experiments in noisy environments. The experimental results indicate that, compared with the baseline performance, the proposed feature compensation method achieved a 12.06 % improvement in accuracy on an overall average. 1