Bone Conduction Combined With Air Conduction To Structure Speech Enhancement System

Air conduction speech is easily disturbed by all kinds of noise in the process of communication, Therefore, signals with the air as the propagation medium naturally are subjected to a variety of noise interference, such as nuisances from channels and other speakers. This paper introduces the bone conduction speech signal, which collects audio source not from sensors, but by collecting the vibration of the skull through the highly sensitive vibration sensor and then converting into an audio signal. The advantage is shielding background noise at the source of the sound. But speech signal collected through the human body is along with the serious high-frequency attenuation so that the bone conduction is not as good as the conventional air conduction device in the high SNR environment. In order to play their bone conduction and air conduction characteristics of information superiority, this paper focuses on building speech enhancement system based on bone conduction and air conduction. The main contents are as follows:Firstly, based on Lab VIEW and NI eight-channel data acquisition card, collect and preserve the simultaneous voice of six passages, respectively the bridge of the nose, forehead, ear bones, throat, lips cheeks and air conduction speech. In order to prevent taking a very longtime to gather a large number of data, buffer overflowing and losing data, acquisition and preservation program in the paper has taken a producer-consumer model, namely while collecting and preservating.Secondly, study a variety of air conduction speech enhancement technology to prepare for the subsequent six-channel data fusion. This article not only use more classical spectral subtraction method, wiener method, as well as improved algorithms for every classic enhancement algorithm, but also use more novel wavelet transform method and subspace method. By conducting the simulation experiments with each method, it can be seen that each method has advantages and disadvantages. The type of noise that can be removed by classical methods are different, so the classical methods need to use different speech enhancement algorithms to abate noise and increase the clarity, which has a few limitations. Relatively speaking, the wavelet transform processing speech has smaller distortion and better denoising effect, similarly as subspace method. However, the SNR range of the above methods commonly is 0d B ~ 15 d B,all above methods are not applicable when the SNR is low to-5d B,.The second aspect is to spectrum spread based on speech signal conducted by bone and its implementation. First, extracting the feature parameters of six way speech, feature parameters of air-conduction speech as target output, and the feature parameters of five way speech signal conducted by bone is used as the input, then based the successfully trained DNN model are applied in convert speech signal conducted by bone to speech signal air-conduction and keep up speech signal air-conduction in order to compensate for speech signal conducted by bone of high frequency to make it more clarity.This article focuses on completing the adaptive integration between the air conduction voice and bone conduction voice after five road spectral broaden, estimating the intensity of the background noise and achieving adaptive increasing the weights of the air conduction voice, weakening the weights of bone conduction voice in the high SNR environment, the contrary is the case in the low SNR environment.speech enhancement effect is better and the practicality is stronger based on his method.