Performing successful research on end-to-end speech processing problems requires the integration of many individual tools (e.g. for data cleaning, acoustic model training, language modeling, data analysis, real-time audio, decoding, parsing, synthesis, etc.). It is difficult for new researchers to get started in the field, simply because a typical lab environment consists of a hodgepodge of tools suited to a particular computing set-ups. This environment is hard to recreate, because few people are experts in the theory and practice of all these fields, and can debug and replicate experiments from scratch.
This research infrastructure project creates a "kitchen" environment based on Virtual Machines (VMs) to promote community sharing of research techniques, and provides solid reference systems as a tool for education, research, and evaluation. We liken VMs to a "kitchen" because they provide an environment into which one can install "appliances" (e.g., toolkits), "recipes" (scripts for creating state-of-the art systems using these tools), and "ingredients" (spoken language data). The kitchen even holds "reference dishes" in the form of complete experiments with baseline runs, log-files, etc., together with all that is needed to recreate and modify them.
The project is developing a community and repository by (a) building pilot VMs, (b) engaging the community in using and continuing to develop them on its own, and (c) evaluating the impact of providing VMs for education and research. We envision researchers as well as students downloading a VM, reproducing the baseline experiment, implementing changes, posting their results in the community, discussing with other users who have worked on the same VM, merging improvements back into the VM, which get re-distributed, and finally publishing easily reproducible results. Work with curriculum and project development will support the creation of engaging activities to specifically encourage students at undergraduate and graduate levels.
