Implementation of Speech Recognition with Floating-Point and LNS Arithmetic in SOPC System

Autor: Heng-Sheng Shen, 沈�琤�
Druh dokumentu: 學位論文 ; thesis
Popis: 93
Speech recognition plays an important role in consumer electronic products. Speech recognition involves a sequence of floating-point addition, subtraction, multiplication, division, and some complex function computation. The computations of these operations increase the time of speech recognition. The logarithm number system (LNS) arithmetic has the advantages of high performance, high precision, and low power consumption in computing complex functions. It can reduce multiplication, division, square, and square root operations into simple addition or subtraction operations. Then, it simplifies the complex computations and reduces the time in speech recognition. However, due to the large hardware cost in LNS addition/subtraction, we use both floating-point and LNS arithmetic in our implementation of speech recognition. We implement the speech recognition system in the Altera’s Nios SOPC system. We developed the software of speech recognition based on the embedded Nios processor and by utilizing the hardware/software development system provided by the Altera’s SOPC development system. We designed the hardware of the IEEE 754 single precision floating-point adder/subtractor, the LNS multiplier/divider, and the circuits for the conversion between the LNS format and floating-point format. These hardware circuits are integrated into the SOPC system by utilizing the custom instruction feature of the Nios SOPC system. These custom instructions can be called by the speech recognition software, and thus the recognition speed can be enhanced. From our experimental results, the proposed method, implementation with floating-point and LNS hardware, has very small amount of degrade in recognition rate. However, it has improved the recognition speed by 93.5% and 54.7%, respectively, than the implementation with pure software and the implementation with floating-point hardware.
Databáze: Networked Digital Library of Theses & Dissertations