Screening of pathological voice from ARS using neural networks

These days there are many attempts to diagnose patient’s voice by only acoustic voice signal. The major purpose of such experiments are screening the possible patients and lead them to the early diagnosis and treatment. [1][2][3][4][5] Acoustic screening of patient's voice is important in terms of many aspects. It is important for both doctors, who is not an expert of vocal diseases and the patient who want to check their voice without going to the clinic. If such a test is done on the public telephone network, more non-expert doctors and more patient can diagnose their voice without going to the hospital. Of course the purpose of this kind of approach is not to give them an exact diagnosis or treatment. It is only to give them the warnings to visit the hospital and to let them find vocal diseases earlier. The background of this kind of research is that most of the patient who come with severe vocal disease are too late when they arrived to the hospital. Even though there are some previous researches about discrimination of pathological voice, those were based on pre-recorded ones from the hospital. So the usage was limited to the patients who visit the hospital. If such kind of diagnostic information is provided through the telephone network, it will be much efficient. But the problem when trying to use the telephone network is that the quality of voice from the telephone is not as good as normal high quality recordings. In previous researches wide-band(25KHz) signal was used. [2][3] So we tried to investigate the change of acoustical characteristics and tried classification based on the result in parallel. In the following sections it is described how the collection, analysis and classification is performed using ARS voice. Also the acoustical differences between ARS voice and DAT voice are measured.