Research On Pitch Detection Techniques At Low-Rate WI Speech Coding

Waveform Interpolation speech coding is one of the most potential low-rate speech coding algorithms in recent years. With high performance, WI technique has been widely concerned. Many research institutes around the world are focusing on and developing WI speech coding algorithm, expecting to give a communication quality of synthesized speech at 2kb/s or even below. This is the background of our WI speech coding project.Pitch detection is one of the most important tasks in low-rate speech coding field. The accuracy of pitch detection will affect the performance of the whole CODEC. In this thesis, the research focuses on pitch detection techniques of the low-rate WI speech coding. Aimed at the problems of voiced-unvoiced error, pitch doubling and halving, accuracy of pitch detection and pitch quantization, a series of pitch detection techniques including pre-processing, pitch detection and pitch quantization were proposed.The original pitch detection algorithm in WI coder is based on Normalized Cross-Correlation Function (NCCF). Because of the high quality in clean speech, NCCF algorithm is enough for WI coder. Based on all these above, the results of main research are as follows:(1) A Voice Activity Detection (VAD) algorithm based on DCT band-partitioning spectral entropy is proposed. As the pitch doubling and halving problems of NCCF algorithm often occurred with varied noises and Signal to Noise Ratio (SNR), VAD algorithm is employed to separate speech and non-speech segments. Because of the large difference between spectrum of speech and noise and great ability of DCT to reduce the correlation of signal, it is advantage to use DCT domain spectral entropy to do VAD. Experimental results in different noises and SNR indicated that this VAD algorithm can divide speech segments from non-speech segments accurately and reduce voiced-unvoiced error obviously.(2) An improved DCT-HN speech decomposition algorithm based on the Harmonic-Noise Model is presented. In order to provide an accurate-pitch-cycle speech for pith detection algorithm with varied noise and SNR, we use signal decomposition theory in pre-processing of pitch detection. The improved DCT-HN algorithm converges quickly and has a little better performance than before. So the SNR of noise-corrupted speech is enhanced in some extent and accuracy of pitch detection in low SNR is assured.