A Classification-Based Non-local Means Adaptive Filtering for Speech Enhancement and Its FPGA Prototype

Non-local mean (NLM) adaptive filtering is a well-explored technique for the denoising of images and electrocardiogram signals. In NLM filtering, the signal value at a particular sample point is estimated by a weighted average of sample points over a search neighborhood. The NLM filter effectively removes the noise when there are similarities among the samples of the signal over the search neighborhood. Due to the time-varying nature of the vocal-tract system and excitation source, the magnitude and frequency of the speech signal vary over the time. Consequently, NLM filtering is not effective in removing the noise components from the speech signal. The similarity among the sample points can be improved by classifying the speech signal into different categories depending on the magnitude and frequency components. In a given speech signal, the vowel-like speech (VLS) are high-magnitude regions compared to the other non-VLS. The vowel, semivowel and diphthong sound units are collectively termed as VLS. In this work, at the first level, the noisy speech signal is classified into VLS and non-VLS for improving similarity. Next, the non-local similarity present within the VLS and the non-VLS is exploited separately for an effective speech enhancement through NLM filtering. The experimental results presented in this study show that the proposed approach provides better denoising performance when compared with the NLM filtering without speech classification as well as recently reported speech enhancement methods. The hardware architecture of the proposed approach is also designed and prototyped on FPGA.