Retrieval And Authentication Technology Based On Speech Perceptual Hashing

With the arrival of "Internet+" era, and the rapid development of network communications, people can communicate directly with the computer or various mobile terminal devices, and complete a variety of voice commands, which not only bring convenience to people but also lead to the storage of voice "explosive" growth. The birth of cloud computing provides a turning point for massive information storage. But cloud storage platform is not a trusted third party, how to ensure the security of data and how to improve the efficiency of large-scale voice data processing in cloud become the urgent problems.To solve these problems, the speech perceptual hashing and its application in the large-scale speech retrieval and authentication are studied in this thesis. The main work is as follows:(1) This paper presents a perceptual hashing algorithm based on formant frequency and energy difference of time-domain. The formant frequency which can reflect the characteristic of speaker’s tone is extracted as rough characteristics, and the energy difference of time-domain is extracted as detail characteristics. The rough and detail characteristics were quantified as the perceptual hashing digest. Rough sequence and detail sequence are combined as the final perceptual hashing sequence. Simulation results show that the proposed algorithm has strong robustness and distinction. The algorithm is designed for application purpose, so when the proposed perceptual hashing scheme is used in large-scale retrieval, it will greatly improve the effectiveness.(2) The feature selection and quantification methods of perceptual hashing algorithm are varied, but the matching process in a variety of applications are usually matched individually, and choose the most relevant match as result, this approach increases the number of unnecessary computation. This paper proposes hierarchical matching idea, which can significantly improve efficiency and provides a new idea for the cloud retrieval application. The matching process of the speech retrieval program is as follows:first, matching the coarse features, filtering the speeches which have a similar timbre of the speaker, and then, matching the detail characteristics of these speeches. Finally, the exact match was found. This will eliminate a lot of unnecessary matching calculation, so that the matching efficiency greatly improved. To ensure the security of data transmission process, the search results will be certified and only the properly certified will be returned to the user. Experimental results show that the retrieval scheme obtain a higher recall ratio and precision ratio, while the retrieval efficiency is also improved significantly.