ELA is non-generative AI-driven learning assessment software for teachers to use in their classrooms.
ELA comprises the following components (with their github repositories linked)
- ELA AI, the AI engine
- ELA Collect, an Android mobile application for teachers to capture learner outputs in the classroom. This is a forked customization of the renowned ODK Collect app.
- ELA Web, a browser-based web application for teachers to control the entire assessment process
- ELA Middleware, a technical component hidden from users that orchestrates interactions between ELA Web and ELA AI to reduce computer resource usage.
A companion component is the ECUBE data collector, a custom data collection tool that extracts learner outputs from the ECUBE English Language Lab (ELL) software developed by KITE. Read more about KITE here and the ECUBE ELL here
- Pedagogy Driven : Designed by educators and technologists at IT for Change, ELA is driven by pedagogy, not technology.
- Teacher First : ELA is teacher-facing software. No learner <-> computer/phone interaction is necessary for the learning process.
- Offline First : ELA is carefully architected to run on a desktop computer. No internet required.
- Privacy Focus : With no internet required, all learner data and assessment reports stay on the computer. No data is shared or visible to any party, including the app developer.
- Openness : ELA is licensed to be free and open source (FOSS). Always.
- 'ELA' is a polysemous acronym. That is, an acronym with multiple, related meanings.
- Language learning assessment is central to the software, and the English language in particular. ELA is therefore 'English Language Assessment'.
- In the immediate roadmap is support for other languages (Indian languages, to begin with). Here, ELA is therefore also 'E Language Assessment', the 'E' stands for 'Electronic' of course.
- A future iteration of the software will also support other forms of assessments based on non-speech inputs such as MCQs that the teacher will record based on learner responses, and OCR scans of paper-based MCQ responses of learners etc. This could be in any subject of learning. Here, ELA is 'E Learning Assessment'. Note the absence of a hyphen between 'E' and 'Learning'. This is not 'E-Learning'. Learning happens in the classroom.
- Lastly, in homage to the Free Software Movement and The GNU Project, ELA is of course also a recursive acronym that expands to ELA Learning Assessment
- Assessment of English language learner's speech, with focus on early speakers.
- Support for monologic and dialogic oral productions (the dialog is between the learner and the teacher)
- A 9-point scale based reporting that is sensitive to early speakers' abilities in their home language(s)
- Recommendations of audio stories for learners to listen to, to improve listening skills, vocabulary and overall proficiency in the language (Work in progress)
To successfully install and run ELA, the following are the minimum requirements.
- A desktop / laptop computer with 8 GB RAM, 4 CPU cores, 50 GB disk space is sufficient. Note: The AI components are resource intensive when they are running. Avoid actively running/using other software when the AI components are executing.
- Any standard Linux OS distribution. While other OSes/versions should work, ELA is developed and tested on Linux, specifically Ubuntu 22.04+
- The desktop should have the following software installed
- Docker CE
- KDE Connect app to transfer files between ELA Web and ELA Collect. An alternative to this software is raw Bluetooth connectivity, but working with Bluetooth connections directly is harder.
- A USB storage device with about 20 GB space for the ELA software, AI models and tutorial materials, and a working USB port on the desktop computer.
- An android mobile phone with Android OS version 10 or above (ensure the phone microphone is in working condition)
ELA uses the following AI and Machine Learning components
- pyannote, for speaker diarization to separate learner and teacher speech in conversations
- whisper, for spoken language identification and transcription of learner speech
- spaCy, for natural language processing of transcribed learner speech
- A custom grammar model built on the lines of vennify, supported by the Errant library
See here for installation instructions, technical design and user guide.