Noise-robust Pitch Detection Algorithm Based on AMDF with Clustering Analysis Picking Peaks

A peak-picking method has been proposed and integrated into baseline AMDFs for pitch detection. Many of the known peak-picking methods for AMDF/ACF based PDAs expect to search the sole peak located at pitch period. However, it is a fact that correct pitch information is also contained in the peaks located at positive integer multiple of pitch period, whose existence are strong evidence of the validity of detected pitch as well. The proposed peak-picking method uses clustering analysis searching these peaks as many as possible. A necessary condition and some constraints are used to select the optimal cluster produced by clustering analysis. The proposed method shows a very good adaptability and noise-robustness. 5 improved AMDFs are employed to evaluate performance of proposed peak-picking method. Gauss white noise is added into speech for anti-noise tests. Experiments indicate slight improvements in low-noise environments and clear improvements in high-noise environments compared with the reference PDAs.