Research On The HNR Separation Of Voice Source With Harmonic Noise

HNR (Harmonic to Noise Ratio) reflects the degree of hoarseness voice to a large extent which is one of the main characters to diagnose throat diseases. It plays a significant role in the field of phonetics and throat disease treatment.In this study, DTW,ICA and wavelet analysis are came into use of analyzing and calculating the HNR of Chinese pronunciation [a:], 10 normal and 10 different pathological changed steady-state vowel signals made by labial tip. It can evaluate the degree of voice hoarseness by the degree of HNR which can be used to distinguish normal voices from pathological changed ones. DTW (Dynamic Time Warping) is a kind of time-domain method which matches waveforms on time and the differences among the rests used as components of noise.When this study tests each voice cycle starting point, it uses auto-correlation instead of minimum value of original test. ICA (Independent Component Analysis) uses linear prediction inverse filtering to get glottis airflow volume velocity wave, then separate harmonic wave from noise in glottis airflow volume velocity wave. The magnitude of outputting harmonic wave and noise signal changes after separation. This study normalizes input and output signals to solve this problem well. The noise caused by vocal cord lesions and injuries is a kind of random noise. Since noise mixing into signal will cause singularity, the size of singularity can be measured by Lipschitz index.Random noise and effective signal have different Lipschitz index of singular points, so their wavelet transform modulus maxima are not the same of different scales. This study uses wavelet analysis to separate harmonic wave from noise in glottal airflow volume velocity wave and calculate the degree of HNR.This study uses three methods above to do the steady-state voice experiments and get some degree of HNR which can distinguish normal voices from pathological changed ones. Independent component analysis is the best method, followed by wavelet analysis and dynamic time warping method is the worst. Finally, it analyzes the advantages and disadvantages of them.