EdulingAI for Languages: Technical and Pedagogical Design and Engineering and Demo
Description
In this live session, Linh Phung and Nikolas Wolfe introduce the latest developments for Eduling, including the expansion of their mobile application into a web-based platform (app.eduling.ai) and the addition of multi-language support spanning 144 target language varieties. The discussion focuses on the intersection of task-based language teaching (TBLT) and speech technology. Nik explains the engineering behind low-latency bidirectional streaming, automatic speech recognition (ASR), and text-to-speech (TTS) integration. They demonstrate live interactive tasks in languages such as Spanish, Vietnamese, Twi, Frafra, and Ga, critically examining the current capabilities and limitations of large language models when handling low-resource, less commonly taught languages (LCTLs).
Summary
Introduction to the Eduling Web App and Platform Updates
Eduling CEO Linh Phung and Eduling CTO Nikolas Wolfe open the session by highlighting a major milestone for Eduling: the transition from a mobile-only application (iOS and Android) to a fully functional web-based endpoint (app.eduling.ai). This web version provides learners with broader accessibility and a more spacious desktop interface. The extra screen real estate allows users to view image-based tasks, diagrams, and side-by-side writing tasks such as in IELTS Writing Task 1 more comfortably while interacting with the multimodal dialogue agent, EdulingAI.
Nik details the extensive engineering effort required to bring the web app to life. Because the system relies on bidirectional streaming for everything, it eliminates the traditional latency associated with waiting for speech recognition to finish or entire text-to-speech files to synthesize. Audio playback begins dynamically while data is being synthesized, creating a responsive, low-latency dialogue environment designed to mimic the communicative pressure and flow of real-life conversations.
The Evolution of Multi-Language Support
While Eduling was originally built to serve English language learners—addressing a massive global demand—the platform has now expanded to support 144 target language varieties. Nik explains that this multi-language functionality was discovered somewhat by accident during backend architectural updates. By designing the system to “play” first languages (L1) and target languages (L2) dynamically, the team realized they could swap speech recognizers, machine translation engines, and text synthesizers to teach and support almost any language.
For Nik, an L1 English speaker, experiencing Eduling in a language he is unfamiliar with provides critical insight into user pain points. He emphasizes that until a developer or designer experiences the rush of a system firing speech at them in an unknown language or script, it is impossible to fully understand what learners need to make input comprehensible.
The Technical Architecture of Spoken Language Interaction
Nik breaks down the three foundational pillars required to support spoken language interactions within a platform like Eduling:
Automatic Speech Recognition (ASR): Converts acoustic features captured by a microphone into a probable sequence of words using an acoustic and language model.
Dialogue Management: Processes the transcript via a large language model backend (such as Gemini) to generate a contextually relevant response.
Text-to-Speech (TTS): Converts the output text back into synthesized voice audio.
Nik highlights that while major commercial providers offer mature technology for high-resource languages (like English, Spanish, French, and Mandarin), the digital services landscape largely gatekeeps less commonly taught languages. To bridge this gap, Eduling integrated specialized regional providers, such as Khaya AI and Ghana NLP, which build robust ASR, translation, and TTS resources for West African languages that are typically ignored by mainstream tech giants.
Live Demonstrations and Task Design Insights
Throughout the session, Linh and Nik walk through live demonstrations of the platform across multiple languages, highlighting both successes and failure modes:
Spanish: Linh demonstrates beginner-level tasks using food-related visuals (tacos) and multimedia video songs (children's daily routines). Linh shares her personal learning strategy: mirroring phrases spoken by the AI to keep interaction moving forward, receive more input, and reduce the anxiety of producing grammatically perfect output.
Twi & Ga: Nik demonstrates tasks in Akan (Twi) and Ga, pointing out how regional dialects (such as Akuapem variants injected by underlying language models) and translation artifacts emerge. They candidly discuss the limitations of foundational models when dealing with low-resource languages, noting that hallucinations and awkward phrasing occur because massive commercial models are predominantly trained on high-resource languages.
Frafra: Experimenting with Frafra, a small language community in northern Ghana, they observe how mismatched speech models and translation engines produce humorous misunderstandings, underscoring the reality that prompt design alone cannot overcome underlying data scarcity in low-resource languages.
Pedagogical Value and Future Collaboration
Linh invites educators and researchers to collaborate with Eduling to experience and experiment with tasks in these languages. She argues that learning a new language is one of the most effective forms of professional development for language educators. By stepping into the shoes of a beginner learner, teachers remind themselves of the second language acquisition process. Furthermore, collaborative task design acts as a form of action research, where educators analyze data, refine the tasks, and iteratively improve learning materials.
The session concludes with an invitation to teachers, researchers, and language enthusiasts. Linh and Nik encourage educators attending the upcoming "Digital Approaches to Less Commonly Taught Languages" conference and the wider community to test the platform, provide critical feedback, and collaborate on building tasks and resources. Interested educators can reach out to the team at info@eduling.org to help shape the future of technology-mediated language learning.
