Multi-Model Noise Suppression using Particle Filtering

We propose a noise suppression method based on multi-model compositions using particle filtering. In real environments, input speech for speech recognition includes many kinds of noise signals. For such noisy speech, we previously proposed Multi-model Noise Suppression (MM-NS) that uses many kinds of noise models and their compositions obtained from training data. However, since MM-NS only uses the static property of noise models, handling unknown noise distributions is difficult. We introduce a particle filter into MM-NS. The distributions of noise models are used as prior distributions of particle filtering to increase the accuracy of the estimation of noise signals for input data. We evaluated this method using the E-Nightingale task, which contains voice memoranda spoken by nurses during actual work at hospitals. The proposed method outperformed the original MM-NS.