关键词: 统计模型, 语音活动检测, 两步降噪技术, 子频带, 决策导向

Abstract: To reduce the influence of noise and the frame delay of Decision-Directed (DD) parameter estimation algorithm on the robustness of Voice Activity Detection (VAD), Two-Step Noise Reduction (TSNR) technique was used to track the priori SNR of the current frame to improve the estimation accuracy of parameters in speech transient. Then, based on the frequency-selective property of noise signals, frequency spectrum was split into sub-bands to isolate noise corruption within particular sub-bands, so that the sub-bands weighting algorithm, in which sub-bands features were combined with their reliability to provide judgment, was constructed. The experiments show that the sub-band weighting method is superior to those full-band ones such as Sohn, Cho and G.729 B.

Key words: statistic model, Voice Activity Detection (VAD), Two-Step Noise Reduction (TSNR), sub-band, Decision-Directed (DD)