Hardware-Software Codesign of a Large Vocabulary Continuous Speech Recognition system.

Abstract

Modern real-time applications with increasing design complexity have revolutionized the embedded design procedure. Energy budget constraints and shortening time to market have led designers to consider cooperative design of hardware and software modules for a given embedded application. In hardware-software codesign the trade offs in both the domains are carefully analyzed and the processor intensive tasks are off-loaded to the hardware to meet the performance criteria while the rest is implemented in software to provide the required features and flexibility. Speech recognition systems used in real time applications involve complex algorithms for faithful recognition. The nature of these tasks restricts the implementation to large platforms and is not feasible to meet the performance constraints for smaller embedded mobile systems and battery operated devices.

This thesis proposes an idea for hardware-software codesign of a Hidden Markov Model (HMM) based large vocabulary continuous speech recognition system. The entire procedure can be divided into three phases: the initial phase deals with the spectral analysis of the speech input, the second phase deals with learning of the sound units followed by the recognition phase. Studies have shown that the recognition phase consumes more than 50% of the processor time. Keeping this in mind, we partitioned our design to perform the spectral analysis and acoustic training in software using the front end executables and the acoustic trainers provided by the CMU SPHINX. The decoder implementing the phonetic detection and viterbi algorithm was designed in hardware. In this project we simulated different speech input files in software and the relevant input vector files required for hardware analysis were tapped from the SPHINX system