The Research And Implementation On GPU-accelerated Audio Retrieval Algorithm

With the development of digital movie technology, the multimedia information experiences a explosive growth. As a component of multimedia, audio information also grows a lot. How to get the interested content from the massive audio data becomes a research focus and audio retrieval has been created. Audio retrieval means that identifying the corresponding audio content from the target audio by analyzing audio characteristics. Audio retrieval method is divided into two categories. One is content-based audio retrieval, which mainly uses audio characteristics for classification and comparison. And the disadvantage of such method is that the technology is more complex while the retrieval accuracy is difficult to cross the ’semantic gap’. The other method is based on the similarity, also known as specific audio retrieval. It does not need to identify the sound and scene, or define and train the model in advantage. And it can directly use audio characteristics to traverse the target audio and get the results. This method is simple and has a high correct rate, but it has the following problem that computing cost and retrieval time are both proportional to the length as the retrieval target.For the problems in traditional audio retrieval method, this paper presents a GPU-accelerated audio retrieval which has improved performance a lot compared with the traditional method. The study mainly focuses on the following aspects:1. This dissertation analyzes the architecture of general programming on GPU and CUDA platform. From the data point of view, it describes the parallel algorithm to speed up the calculation, and maps the parallel algorithm to the GPU through CUDA platform;2. Extract audio characteristics. In order to fully express audio characteristics, we use the audio time domain feature, audio energy and zero-crossing rate feature; and audio frequency domain feature, Mel cepstral coefficients. In order to speed up the calculation, this paper takes the parallel algorithm based on CUDA architecture. Experimental results show that compared to the CPU algorithm, the speed up of feature extraction on GPU can be more than 12 times;3. This paper proposes a GPU-accelerated audio retrieval method. Firstly, the audio is divided into multiple short audio segments based on the features, and the eigenvalues sequence is calculated from each short audio segment using the GPU-accelerated algorithm. Then use the suffix array deformation algorithm to find the common set from two eigenvalues sequence. Refine and overall match algorithm is used to get the retrieval result from the common set. Experimental results show that the retrieval accuracy is over 90% and compared with existing algorithms, this method can significantly improve the retrieval speed and speedup can be achieved in more than 10 times. In summary, this paper summaries the advantages and disadvantages of the GPU-accelerated audio retrieval algorithm and gives a prospect on GPU-accelerated audio retrieval algorithm development.