A novel instantaneous frequency-based voice activity detection for strong noisy speech

The development of robust voice activity detection (VAD) for strong noisy speech is a challenging task. In this paper, we propose a novel voice activity detection method under Hilbert-Huang Transform (HHT) framework by using its good ability to automatically extract signal-frequency related intrinsic mode functions (IMF) by the empirical mode decomposition (EMD), which provides us a more flexible way to select different IMFs with strong speech components. With the Hilbert transform, the instantaneous frequency (IF) can be computed. Making use of the speech characteristics of each IMF, a weighted instantaneous frequency average (WIFA) measurement is proposed and the corresponding WIFA-VAD algorithm is developed, where the VAD threshold can be automatically estimated using the first noise frames. Experiments show that the proposed WIFA-VAD can achieve comparable results at high SNR. For low SNR (e.g., -5dB and below) conditions, it is able to achieve lower false alarm ration (FAR) and missing error ratio (MER), compared with those of the conventional VAD algorithms.