Empowering Autonomous Learning Through AI

I recently developed a tool that gave birth to the first interpreting lesson entirely generated by Artificial Intelligence. The material (example available here) showcases the potential of Generative AI to transform teaching methods across various subjects, placing students at the heart of the learning experience by making them more autonomous.

The dream of shifting the learning process from being teacher-centric to student-centric resonates with many. This shift aims to liberate students from their reliance on trainers, empowering them to learn in ways that suit their individual preferences for timing, location, and didactic needs. In this envisioned model, trainers transform into guides who facilitate learning rather than simply dispensing knowledge. This concept, including its application in the fields of translation¹ and interpretation² , has been a subject of theoretical discussion for some time. Over recent decades, significant changes have occurred in teaching and learning methodologies, and technology promises to drive this evolution even further.

Consider the specific case of training conference interpreters. It’s widely acknowledged that developing the necessary skills for effective interpretation requires extensive practice. This training involves progressive and guided practice sessions, which relies on the availability of the right teaching exercises (speeches and the like). However, sourcing appropriate materials for interpretation practice can be challenging. Learners often need materials tailored to specific needs, such as speeches with a particular level of difficulty, topic, structure, or style. They may require speeches that are rich in numbers or figurative language, resemble the style of certain speakers, fit specific contexts, or vary in formality and clarity of delivery. Despite the abundance of speeches available online and in curated repositories like the Speech Repository, finding materials that offer the necessary level of customization and detail to meet individual or class-specific needs is not always feasible. The result: a gap between what learners use to practice (or are given to practice) and what they would really need to improve.

Things might be about to change. Generative AI has advanced to a level where it can produce training materials of exceptional quality, offering – and this is the main point – a wide range of customization options. The possibility to have a wealth of speeches fine-tuned to training needs offers a unique opportunity to empower learners (but also trainers).

To experiment with this opportunities, I have created a simple pipeline that combines Large Language Models (LLMs) for the creation of speeches and related activities with text-to-speech synthesis for audio delivery. All packed together in a simple to use application. The LLM component includes a series of prompts designed to generate speeches with a variety of variables, such as: length, language, domain, lexical complexity, structure (in a spectrum from well-organized to bad-organized), register, richness of numbers, terminology and proper names, etc.³ The creation of the right prompts that can be used in a click-and-play modality is quite a time-consuming activity, and in my case I have limited the scope to the speech formats I need to create right now for my teaching purposes.

Given a super short series of parameters by the user, the tool produces, full automatically, the following:

1. audio of a speech (in quite an astonishing quality)
2. script of a speech for evaluation (to be honest nothing different to real speeches like the ones in the Speech Repository)
3. list of key terms for learner preparation
4. short info about the speech for learner preparation
5. comprehension questions for post interpretation activities
6. terminology and lexical activities for post interpretation activities

In my demo, all this info is packed into a user-friendly webpage like the one linked here. Of course everybody can produce similar materials having access to a LLM and a good TTS engine. A bit of time to experiment is also needed. But having everything packed in a deployable app is of course the way to go.

I recently incorporated the didactic unit in the example linked above into my university class. Interestingly, the students did not realize that the exercise was AI-created until I revealed it at the end of the lesson (I guess they would have if they were confronted with it without the need to interpret). The exercise proved to be more challenging than anticipated.

A side note: I developed this application for personal use. My plan is to use it to create specific exercises for my students, foremost for self-learning purposes. I believe it is particularly handy when they need to train specific aspects. Ideally, however, this or a similar app should be made directly accessible to them, as it would significantly empower their learning process. Imagine the impact of having a high-quality speech generator just one click away for them. This leads me to wonder why universities aren’t more invested in creating such tools. It is easy. I have done it in a raining November morning. Quality, I think, is okay and is certainly about to improve (video is only a step away). Why is there a lack of interest in harnessing these technologies for educational advancement? If academics focused not only on discussing, writing, and applying for grants, but also on leveraging AI to create practical tools for their students, the impact could be significant. This shift from merely debating AI misalignment and similar issues to actively developing beneficial applications could greatly enhance the educational experience.

1. See the (life)work of Kiraly, for example here. ↩︎
2. See for example Fantinuoli/Prandi here. ↩︎
3. See Setton and Dawrant’s Speech Difficulty Index for a proposal of a metric to assess training and exam materials (in Conference Interpreting: A Trainer’s Guide by R. Setton and A. Dawrant) ↩︎