ELPIS

Scope

Recent technical advancements in Machine Learning have lead to significant improvements in Automatic Speech Recognition (ASR). However, due to the technical expertise required, these advancements have mostly been inaccessible to linguists in the field. The ELPIS project was created to address this issue, with the goal to give linguists access to modern Machine Learning transcription tools without needing to be a computer scientist.

The initial version of ELPIS was built over a couple of years by some incredibly smart research assistants, but without unifying oversight. This lack of cohesion in the architecture meant that after the project changed hands a few times and grew in scope, it became unwieldy to work with.

Persephia was contracted under the ARC Centre of Excellence for the Dynamics of Language (CoEDL) to help maintain and add features to the project, and then subsequently to rebuild ELPIS with a view to making ongoing development more sustainable in the future.

Approach

1. In consultation with linguists who were already using or had interest in ELPIS, common pain points were identified.
2. On the other side, as maintainers of ELPIS, we were able to identify common issues we were having during development.
3. A new architecture plan was put in place to address the issues of all parties, and to facilitate use cases which were out of scope of the original software.
4. Key development milestones were identified, and subsequently reached.

Outcomes

1. The complexity of the project was reduced to the stage where new hires were able to be brought in without significant onboarding.
2. The new architecture allowed for cutting edge ASR models to be supported.
3. The core ELPIS library was extracted and reached 100% test coverage.
4. The final stage of the project (still ongoing) is to create a desktop electron app to further improve the installation process.