Voice Pathologic Feature Extraction And Analysis

Compared to the existing diagnostic techniques, the voice diagnostic technology has received sustained attentions in recent years due to its non-pain and non-invasive attributes. In the current literature, the sampling equipments and the sampling processes of the voice collecting vary greatly. However, the relationships between these sampling factors and the diagnosis results have been rarely studied. In fact, the development of the voice diagnosis technology is still in an initial stage, and it is urgent to standardize process of voice collection. This dissertation focuses on the standardization of the collection processes as well as the pathological feature extraction and analysis.To build an objective sampling environment, a voice diagnosis system for collecting voice samples, both healthy and the pathological ones, is designed. Three key components in this system are the sound proof room, microphone and the sound card. The sound proof room is included to reduce potential noise effects and the reduction amount is about 40 dB. In order to ensure that the recorded voice signal is not distorted, the selection criteria for the microphone and sound card have been given. In order to guarantee the objectivity of the sampling process, this dissertation has determined the collection process specification by checking the existing experimental collection process. Then we have determined the voice materials in our collection, according to the characteristics of Mandarin and the specialty of the subjects. The voice materials cover all the vowels of Mandarin and most of its consonants. By this way, we can capture the pathological features in voice caused by the illness of vocal organs.After the design of the collection process, we have used this system to collect 2828 samples at 192 kHz from 101 persons who are healthy person, Parkinson’s patients, lung cancer patients and patients of vocal cord diseases in Guangdong Province Traditional Chinese Medical Hospital. Using these data, we have investigated the differences of estimated pitch and GCI sequence at different sampling rates. Besides, the classifications using some independent diagnostic features and their fusion features at different sampling frequencies are also researched.Experiments show that for Parkinson’s disease, when the sampling frequency is 16 kHz and above, there are only small differences among the contours at different sampling frequency. So are the GCI contours. When using fusion features to classify the voices of diseases and healthy ones, the classification accuracy of Parkinson’s samples is drop significantly at 8 kHz and the classification accuracy of lung cancer samples or vocal cords diseases samples are drop significantly at 8 kHz and 16 kHz, while the difference of classification accuracy of each classification tests at other sampling frequency is about 3%. The best accuracies of the three classification tests are 89%±3%, 87.50%±2.2% and 84.23%±2.9% respectively. With the storage space, the speed for analysis and the accuracy of classification in consideration, we conclude that 24 kHz is the most practical sampling frequency for all of the three diseases. These diseases are typical for the nervous system, lung, and the vocal cords. Therefore, we derived that the conclusion is applicable to other diseases of the three organs.