Ama Speak
A mobile app that pairs English learners with AI companions for speaking practice, delivering personalized feedback grounded in language learning theory.
A mobile app that pairs English learners with AI companions for speaking practice, delivering personalized feedback grounded in language learning theory.
English learners who want to improve their speaking face two compounding problems: finding opportunities to practice, and getting meaningful feedback when they do. For high-stakes test-takers preparing for TOEFL or IELTS, this is especially acute. Existing tools either provide no feedback at all, or feedback so generic it doesn't help learners understand what to improve.
Ama Speak was built to address this with AI-powered tutors that evaluate spoken responses against real assessment criteria and give learners concrete, actionable feedback.
Research
I conducted research on how speaking proficiency is evaluated across established frameworks, focusing on pronunciation, vocabulary, sentence structure, and complexity. From this, I developed an assessment model that integrates metrics including speaking rate, confidence, accuracy, and lexical complexity to build a well-rounded learner profile.
The assessment model informed how the AI tutor evaluates a learner's response and decides what feedback to give. Rather than scoring globally, the model surfaces specific issues, such as overused sentence structures or mispronounced words, and links them to concrete improvement steps.
Content Design
I generated question banks and writing prompts categorized into Beginner, Intermediate, and Advanced levels to support varied learning paces and goals. I also designed teaching steps for vocabulary learning and created a prototype for TOEFL speaking task feedback, aligned with official TOEFL scoring rubrics.
Tutor Design
I designed the chatbot prompts and conversation flows for multiple tutor personas, including career coaches, TOEFL experts, and general speaking companions. Each persona was tailored to a different learner goal. The prompts were carefully written to simulate natural, engaging dialogue while keeping the AI focused on language development objectives.
The tutor-based conversation system was one of the most language-learning-theory-heavy parts of the design. I drew on error correction research and input hypothesis to decide when and how the AI should intervene with feedback during a conversation, rather than waiting until the end.
We didn't anticipate how many learners would want to use the app specifically for TOEFL preparation. This unexpected demand led us to develop a dedicated TOEFL tutor aligned to the official TOEFL scoring rubric, and later an IELTS equivalent. It shifted the product's identity from casual conversation practice toward structured, high-stakes test preparation, which gave the product a more focused value proposition than we started with.
UX Design
I designed the onboarding flow to establish a learner profile upfront, enabling personalized tutor recommendations from the first session. I also designed the Figma mockup for the Call Summary page, focusing on presenting AI-generated feedback in a clear, encouraging, and actionable format.
Ama Speak was published on the App Store and has since been archived. The TOEFL tutor became one of the most-used features, validating the pivot toward structured test preparation feedback.
The product's biggest unresolved challenge is retention. Right now, using Ama Speak feels like phoning a friend. There's no consistent reason to open the app again tomorrow. Unlike a phrase builder or a streak, talking to an AI persona requires a level of intrinsic motivation that most users don't sustain on their own.
If I were building it again, I would design retention mechanics from the ground up: goal-setting during onboarding, progress tracking tied to specific TOEFL or IELTS skill areas, and scheduled practice reminders linked to a learner's test date. The speaking practice product is compelling. The motivation layer is where the real design work remains.
Want to see the product?
The app has been archived.