Research On Special Audio Real-time Retrieval Algorithm And Its Implementation In The Fixed-point DSP

Special audio retrieval, means to retrieve the same or similar audio files as a given queryaudio segment in the audio template library. As a branch of the audio retrieval, special audioretrieval technology has been widely used in press advertising regulation, copyright protection,and garbage voice detection in telecommunications networks et al. Although compared withsome other speech recognition branches, such as speaker recognition, keyword recognition,language recognition et al, special audio retrieval technology is relatively mature. However, inpractical applications, there are still many problems, such as the effect of retrieval precision inthe sub-frame processing when the endpoint is difficult to align, the processing performanceunder the condition of large template library, et al.This paper studied on special audio real-time retrieval in the condition of large templatelibrary (the number of templates larger than3000), based on consistent of some theme projectsin information and technology industry of National High Technology Research and DevelopmentProgram863. The main achievements of this dissertation are outlined as follows:1. A speech feature extraction algorithm based on the adaptive weighted-sum and variableframe rate (AWS_VFR) is put forward. Firstly, the technology of adaptive weighted-sum is usedto generate a reference frame, and then, frame rate is adjusted by the dynamic nature of speechspectrum: taking more frames in the segments that change rapidly in voice characteristics toensure extracting enough available information, and taking less frames in the segments thatchange slowly to ensure the speed of match later. The simulation results show that, the speechfeatures extracted by this method can preferably solve the problem of missing retrieval, causedby that the sub-frame starting point of query audio and audio templates is difficult to align in theprocess of sliding match.2. A special audio fast retrieval method based on template-subspace is proposed. Firstly,according to the characteristics of a certain correlation among the same or similar templates, thetemplate library is divided into a number of template-subspaces, and then taking the two-stageretrieval method, first to find the subspace that belonged to the tested audio through a coarsepositioning process, and then to retrieve all the templates in the candidate subspace precisely tocomplete the pattern recognition, and thus to enhance the retrieval speed. Simulation resultsshow that, this method can significantly reduce the amount of template match calculation toimprove the retrieval speed, in the case of almost no loss of precision, and effectively solve theproblem of the existing retrieval algorithms, that the retrieval time is too long when the storagecapacity of the template library is larger.3. According to the structural characteristics of the fixed-point DSP chip TMS320C6455,the algorithms above were transformed and optimized by floating-point to fixed-pointtransformation, parallel and so on, that achieved16channels’ parallel processing on a single chip,in the case of that the number of the templates is3325. In the specific implementation, thespecial audio retrieval system implemented in this paper is able to complete the task of real-time detection of64channels’ audio data in a model, which contained four TMS320C6455chips,based on the daughter card DP64+exploited independently by the project team, and can achievethe applicational requirement of real-time and accuracy.