SPEECH DECODING OPTIMIZATION STRATEGY WITH VITERBI ALGORITHM ON GPU

• Main Author: RADITYA ARSADJAJA NIM : 13514088, ALFONSUS
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/25353
• Institution: Institut Teknologi Bandung

<p align="justify">Automatic speech recognition (ASR) is a popular thing in the present, but the current speech recognition algorithm is slow and requires a way to recognize sounds faster. One way to speed it up is to use GPU. GPU provides high performance by processing many concurrent instructions in parallel. <br /> <br /> Creating an automated speech recognition system in parallel is not easy, and this final project discusses how to implement automatic speech recognition in parallel, which built in several steps. First, change the data structures to be compatible with GPU, then made some GPU kernels so it is equivalent to flow of the serial algorithm on the CPU. <br /> <br /> This final project also discusses some optimization strategies that can speed up the speech recognition process much faster after getting the correct GPU program. Those strategies are based on the results of profiling and analysis of program execution. The precomputing strategy for all the required acoustic scores results in the most significant optimization. Implementation with a good strategy can generate a speedup of about 5.59 - 6.18x from a CPU implementation.<p align="justify">