Research And Implementation Of Anti-packet-loss Speech Encoding Based On Compressed Sensing

With the rapid development of wireless communication technology, mobile phones and other wireless communication devices are more and more deep into people’s lives. However in some areas that are not covered by mobile operators, people need a method that does not rely on wireless infrastructures and can communicate directly between devices. Although traditional walkie-talkie can meet the demand, it needs qualified frequency band and dedicated devices, which makes it inconvenient.In this paper, we realize a device-to-device wireless VoIP (Voice over Internet Protocol) system under Android programming framework with the help of Wi-Fi Direct protocol. The system is portable, immediately available and easy to use. It utilizes public ISM (Industry, Science, Medicine) frequency band but not qualified frequency band. The system only requires a smart phone that everyone has nowadays and does not require dedicated devices.In applications of VoIP especially wireless VoIP application, packet loss is a serious problem that reduces voice communication quality. Packet loss can lead to low speech quality or long communication delay. Packet loss handling is the means to eliminate or reduce the bad impact of packet loss. Packet loss handling includes packet loss recovery based on senders and packet loss concealment based on receivers. Packet loss recovery such as FEC (Forward Error Correction) usually adds delay to communication or raises bandwidth requirements. While packet loss concealment such as waveform replacement and time-domain modification can only "conceal" packet loss but not recover it.Compressed Sensing is a new sampling theory, revolutionizing the traditional sampling theory based on Nyquist sampling theorem, making the number of sampling points smaller according to the sparsity of signal. This paper proposes a new packet loss recovery technique based on the sender with the idea of Compressed Sensing. The technique described herein bases on the compressed sensing theory, using the short-term stationary nature of the voice signal and the stability of the compressed sensing reconstruction algorithm, the encoded voice signal can be recovered even if some data are lost. The experiments show that using compressed sensing-based anti-dropping speech coding under different packet loss level can make the voice quality measured in PESQ increase at least0.4. There are some detail issues in the process of applying the encoding based on compressed sensing, for example, how to allocate encoded data into packets, how to reduce the bad impact of the sparsity of speech signals being approximate but not strict, how to improve CS reconstruction algorithms to make the decoding process more efficient, how to design a protocol to adapt to the CS encoding, etc. In this paper, these detail issues are discussed in depth, and corresponding solutions are proposed and verified by experiments.