GSoC 2018 Conclusion: Speech Recognition Models for Flux Model Zoo

This article concludes a Google Summer of Code 2018 project focused on integrating speech recognition capabilities into the Flux Model Zoo using the Julia programming language. The author details the development of two distinct neural network models: a modern Connectionist Temporal Classification (CTC) network utilizing convolutional layers, and a traditional framewise recognition model serving as a baseline. A significant portion of the work involved optimizing computational efficiency, particularly by porting Baidu's GPU-based warp-ctc implementation to ensure viable training performance. The post addresses the inherent challenges in speech recognition accuracy compared to image recognition and provides a tutorial for running the newly created models. By contributing these freely available resources, the project aims to bridge the documentation gap in speech recognition systems within the Julia community. The author expresses gratitude to mentor Mike Innes and highlights the educational value of learning GPU programming and neural network architecture, marking the successful completion of the summer coding initiative.