The Cough Signal Spotting In Continuous Speech

Cough is a common symptom of many respiratory diseases. The changing of the cough characteristics will provide important clues in research of the respiration syndrome. Different organizational structures can produce different cough patterns, so different cough patterns can reveal the different pathological mechanisms of cough. The evaluation of cough's intensity and frequency of occurrence could provide valuable clinical information in the assessment of patients with cough.This evaluation of cough severity has so far relied mainly on subjective measures, such as cough reflex sensitivity, assessed by cough visual analogue scores, quality of life questionnaires, cough symptom scores and patient's diaries and so on. The manual examination of cough monitoring recordings is a slow and tedious process, and also it can be mistaken by subjective factors. Meanwhile, the patient's perception may be incomplete and unprofessional. Thus, research on the intelligent cough monitoring system and its recognition algorithms become necessary.To implement a practical, user-oriented cough identification system, approaches to the cough signal spotting in unconstrained continuous speech have been studied based on the acoustic characteristics of cough. In real-word, complete cough signal without interfering from noise doesn't exist in unconstrained continuous speech, which always attended by speech or noise. Thus, finding a solution to detect cough sounds, to obtain the start-end point of cough sound from continuous audio recordings efficiently with higher detection rate and better robustness, is the research pivot.This paper investigates present research situation of cough spotting at home and abroad through referring to literature, compares the differences and identifies the relationship between cough spotting and keyword spotting, and studies the mechanism of cough sound generated by respiratory system and acoustic characteristics of cough. Based on acoustic characteristics of cough, using the short time zero crossing ratio (ZCR) identifies the suspicious coughs in order to get the threshold of short time energy. Then, the short time energy is combined with short time ZCR to implement the endpoint detection of cough in continuous speech. The way of setting short time energy threshold is robust, in addition, the endpoint detection program can remove most speech and noise while maintain a lower undetected error rate. Considering the efficiency of endpoint detection, getting rid of noise and speech from continuous speech stream is impractical, so the classification between cough and interference is needed. In this paper we propose the use of hidden Markov models (HMMs), creating cough model and interference model, to automatically classify cough sounds from all the endpoint detection result. In addition, considering the management of the sampling database play a vital role in training the statistical model and improving the system's identification rate, the building and management of sampling database is discussed.Based on above research results, the cough signal spotting system in unconstrained continuous speech is built on vs2008 platform. The system will help the doctor to know about the progress of the patient's therapy with the identification result. In addition, further development of tussis diagnosis system is practical, and will advance the construction of medical treatment information system. To evaluate the performance of the system, 12 patient's cough signal is recorded in sickroom environment; first, the virtual number of coughs in each recording was identified by two experienced doctors using the graphical user interface (GUI). Second, the recordings were analyzed by automatic cough spotting system. Finally, the comparison between these two results was made to show that: for cough signal, using the endpoint detection algorithm, the undetected error rate is 3.09% and insertion rate is 44.7%, using HMMs as the identification algorithm, the identification rate is 96%, the error rate is 6.2%.