Biography
Enrollment Date: 2010
Graduation Date：2013
Degree：M.S.
Defense Date：2013.05.20
Advisors：Dongmei Li Liyuan Liu
Department：Institute of Microelectronics，Tsinghua University
Title of Dissertation/Thesis：Design and Implementaion of a kind of Multi-mode and Real-time Gammatone Filter Bank
Abstract:
Subband filter bank is widely applied in voice processing system (VPS), and has a significant impact on VPS’s performance. Gammatone filter bank is a kind of auditory filter bank which can describe human auditory characteristics, and was immediately used in the research of VPS after proposed. It was verified that Gammatone filter bank can improve the performance of VPS. However, the high computational complexity makes it difficult to be realized in PC and DSP real-timely. Besides, the existing synthesis algorithm is non-real-time as well. Consequently, Gammatone filter bank has been limited in practical application. In this paper, the Gammatone filter bank was implemented in IIR filter form, whose differential equation was derived from its impulse response. Floating-point simulation model of Gammatone filter bank was built in Matlab with subband separated according to equivalent rectangular bandwidth (ERB) rule. Base it on, fixed-point simulation model was built in Matlab, and word length of the fixed coefficient was determined according the quantization error between fixed and floating point model. To solve the problem that the existing synthesis algorithm cannot be implement real-timely, a low complexity synthesis algorithm was proposed and its real-time characteristic is achieved by compensating filters’ delay. This synthesis can make the whole system real-time and reconfigurable. A cascade folding computation unit structure is proposed for FPGA design. This structure reduced the consumption of multiplier and adder. The whole system can work at multi-mode way which has different number of sub-bands, such as 16, 32, 64 and 128 sub-bands. This make it possible to balance system between accuracy and power consumption. The control unit enable the system reusing the computation unit is designed. FPGA can support the high computational complexity of this design. Besides, IO connected to Speech Enhancement System (SES) was designed to exchange data real-timely making it possible to apply Gammatone filter bank to practical SES. Finally, a Gammatone filter bank system (GTFB) can work in four modes and process voice signal real-timely was implemented in FPGA. It can communicate with SES real-timely well. Verified, it can reconstruct input signal accurately with PESQ 4.3 between input and output signal. Moreover, the system delay measured in test is about 20ms, which can meet system’s real-time requirement well. The performance of SES is improved after applying the output data of GTFB. This design of GTFB promotes the practical process of Gammatone filter bank.