A Research On FPGA Implementation Of Gaussian Mixture Model

Gaussian Mixture Model(GMM) is a probability statistical model, it can describe the speech data acoustic characteristic distribution very well, so in all sorts of speech recognition system it is widely used. At present, language recognition system based on GMM is generally implemented in the general processor platform, which real-time rate is low, while real-time problem of multiple concurrent applications can be solved very well by FPGA.This paper relies on a key project of the National "863" programme, which mainly studies on the FPGA implementation of GMM in language recognition system. According to the key problems of Gaussian Mixture Model FPGA implementation, this paper emphasizes the study of the hardware architecture ADDLOG function and exponentiation computation FPGA implementation method. Finally the FPGA design and implementation of the whole Gaussian Mixture Model module is fulfilled.The main work and achievements of this dissertation are outlined as follows:1. A fixed-point implementation method of posterior probability based on ADDLOG function was designed. By introducing ADDLOG function, a fixed-point implementation of GMM posteriori probability calculation is resolved;2. A polynomial approximation algorithm of subsection-displaced is put forward, by which the fast operation of ADDLOG function is realized. By comparing the relationship between polynomial degree and the size of segments, an implementation method of subsection–displaced is selected. Rapid ADDLOG operations are reached under the condition of precision ensured;3. A FPGA implementation method of combination-shifted exponentiation computation is proposed. By the selection and combination of sequence elements for index domain, the deductions shift was mapped to realize exponentiation computation, and the accurate exponentiation computation is realized;4. GMM module of language recognition system is designed and implemented by adopting pipelining and parallel processing structure and combining the algorithms above. Simulation testing indicated that the performance of multi-access and real-time is satisfied when this model is applied to the language recognition system.