VoIP Voice Auditing Based On Speaker Recognition In Gigabit High-speed Network Environment

VoIP(Voice Over Internet Protocol) is a hot technology in the area of Internet application. VoIP has a low-cost, openness, flexibility and bring great convenience to our lives, but it also bring security problems such as cyber-crime issues. Our subject is to solve security problem by using speech recognition technology to automatically identify the speaker's identity. The content this paper contains:Zero-copy packet capture in gigabit network environment: the basic idea of zero-copy is that NIC sends data to user application through the actual physical memory, shorten the data transmitting path, reduce the number of data copy and system calls, avoid the system switch status frequently between core state and users status. Thus in the process of receiving each packet, the platform can reduce the involvement of the CPU, and save resources effectively. Experimental data show that high-performance zero-copy packets capture platform can always reach wire-speed. Compared to Traditional packet capture platform, The performance and efficiency have been greatly enhancedVoIP network packet analysis and restore: The core idea of VoIP is to sample and encode signal, then transmit it through the IP network transmission; In the receiving end, parse received packets and restore them into voice data. SIP protocol is a signaling protocol for creating, modifying, the releasing of the sessions; SDP(Session Description Protocol) protocol is used to describe the session information and media information; RTP(Real-time Transport Protocol) protocol is a media-streaming protocol, it is responsible to packet the data and realize real-time transmission.In the implementation phase, the first step is to analyze the SIP protocol, access to voice transmission port, voice compression format, and some other information, and then analyze RTP packet that load the voice data, finally, restore the voice data to a wave file.Speaker recognition: the key of speaker recognition is to establish model for each speech segment, the performance of the model is closely related to the selection of feature, speaker recognition mainly uses two kinds of feature parameters: LPCC parameters and Mel cepstral coefficients (MFCC ) parameters, the former is based on pronunciation modeling, the main reaction is the channel feature, the later is established mainly based on feature of the human ear's hearing, make use of the logarithmic relationship between human ear's auditory sensitivity and frequency of the sound. The system uses the MFCC parameters to establish vector quantization model. Vector quantization method is a simple and efficient method, which quantifies the data in the multidimensional space whole and compress the data without loss much information. The key of vector quantization method is carve up the region boundary and identify the comparison measure of vectors. This module is divided into training and recognition phases. In the training phase, the system extracts feature parameters from the speech signal for each speaker modeling. In the recognition phrase, the system first extracts feature parameters from the test speech and establishes a model, and then compares the model to training models; finally get the judgment result by some discriminate algorithms.