Noisy Speech Segmentation Using Non-Linear Observation Switching State Space Model and Unscented Kalman Filtering

A reliable speech segmentation in noisy environments is desirable for segment-based speech enhancement and efficient coding. Switching state space model with hidden dynamics has been shown to lend itself naturally to the speech segmentation problem. However, when noise is present, the distorted observation features lead to a poor recognition and segmentation performance. In this paper, the unscented Kalman filtering (UKF) is used during inference to compensate nonlinearly for the effect of noise on the observed features in the log-frequency domain. The proposed algorithms resulted in a much improved segmentation performance in a variety of noises