Chinese Tone Analysis Based On Acceleration Sensor

Tone is an important feature of Chinese pronunciation,plays an important role in the discrimination of Chinese,the lack of Chinese tone or distortion will seriously affect Chinese language communication fluency and accuracy,affect the social life and production practice,communication and other activities in all aspects,tone is very important for speech recognition,speech synthesis and speech rehabilitation.But the absence of tone is widely existence in many scene,such as total laryngectomy,lack of tone in lipreading recognition,whisper tone acquisition,the tone loss affected the promotion of communication between people and new technology.The acceleration sensor is the device to measure the accelera tion.Compared with the remote sensing,it is measured by its own motion.In this study,mainly used to obtain the tone information,its advantage is the external acoustic noise,high SNR,and compared with traditional acoustic measurement methods(e.g.microphone sensor),it can directly get throat motion information,compared with EMG sensor,acceleration sensor and the wear would not be affected by the skin impedance.By capturing the motion information of laryngeal muscles,we can extract and recognize Chinese tone information on the basis of voice information.The main contents can be divided into three parts: hardware system design,data acquisition and recognition algorithm design.1.Hardware: built in acceleration signal acquisition platform,using the acceleration sensor MPU6050 C produced by invensense company,by acquiring the acceleration information of single axis,in order to ensure the acquisition system in power consumption and volume to satisfy wearable requirements,at the same time,obtain the uniaxial data effectively reduces computational complexity and reduce time of feature extraction and classification;2.Data collection: This is the first time to achieve the acceleration signal Chinese single syllable database structures,the research is the first case,the amount of data up to 24000 times of 40 people,and it will be continued,like continuous statements and complicated voice like whisper;3.Algorithm;based on RMS(RMS)detection algorithm threshold segmentation can automatic extract tone information,feature extraction algorithm is stable wavelet transform(SWT),S transform(ST),short time Fourier transform(STFT),wavelet packet decomposition(WPT)and the time-frequency domain method(TAF),finally with the help of the support vector machine(SVM)classification,the extraction of Chinese four tone using acceleration information has been succeed.The experimental results show that:1.Five feature extraction methods which have never been used for tone extraction and the application have different performance in tone extraction,in which STFT has the characteristics of short time and high recognition rate,which is different from other acceleration signals.2.Feature sorting method can improve feature effectiveness,avoid feature redundancy,make operation efficiency.The highest recognition rate in STFT can be 92.10%.3.By comparing the recognition rate of four different cervical muscles,the best location is between the hyoid bone and the thyroid cartilage;4.The five feature extraction method in four sound tone recognition,and we found that the first sound recognition rate is the highest,and the recognition is more stable.