Accent Guesser

Inspiration

Accents carry so much information—where we learned English, what sounds we’re used to, and which parts of pronunciation still feel “foreign” to our mouths. I wanted to build something that’s both fun and genuinely useful: a lightweight way to hear what your English sounds like to others, and a practical starting point for improving clarity without turning language learning into a chore.

What it does

Accent Guesser (AccentGuess) lets you record a short English sample and instantly get an AI-based accent guess. Beyond the label, it’s designed to highlight the pronunciation patterns that influenced the result and offer simple guidance you can apply right away. It also supports a quiz-style flow where you read multiple sentences so the system can make a more confident call and return a richer “accent profile.”

How I built it

I built it as a simple pipeline optimized for speed and consistency:

• A clean browser-based recorder with clear instructions and guardrails (so users know exactly what to read and how long to speak).
• Audio preprocessing to normalize volume and reduce noise sensitivity, then feature extraction focused on phonetics and prosody (vowels, consonants, rhythm, stress).
• A model layer that scores likely accent families/regions and produces confidence estimates.
• A post-processing layer that turns raw model signals into a user-friendly explanation plus targeted practice tips (what to work on, not just what you “are”).
• A results UI designed for sharing, with an optional multi-sentence quiz flow to improve stability of the output.

Challenges I ran into

• Similar accents are genuinely hard to separate. Some accent patterns overlap heavily, especially with non-native speakers influenced by multiple languages.
• Recording quality varies wildly. Mic differences, background noise, and speaking speed can shift results, so the system needed robust normalization and clear user instructions.
• Balancing honesty and friendliness. The product has to be transparent about uncertainty without making results feel meaningless.
• Making feedback actionable. “You sound like X” is entertaining, but real value comes from practical guidance users can actually practice.

Accomplishments that I'm proud of

• A “low-friction” experience: record → get results fast, without complicated setup.
• Results that feel human: a friendly explanation and improvement suggestions instead of a cold classification.
• A quiz flow that encourages better samples and produces more stable results.
• A product that can serve both casual curiosity and serious learners without changing the core UX.

What I learned

• The best ML UX is often about reducing variability (controlled prompts, consistent recording flow) rather than chasing complexity.
• Users trust results more when you present confidence and clear next steps, not just a label.
• Pronunciation feedback is as much product design as it is modeling—how you explain matters as much as what you predict.
• “Fun” can be the gateway to learning when the experience stays lightweight and non-judgmental.

What's next for Accent Guesser

• More actionable coaching: short practice drills tailored to the user’s top two pronunciation challenges.
• Better robustness: clearer mic checks, noise detection, and guidance for getting a clean sample.
• Finer-grained results: more regional nuance where the model is confident, and clearer grouping when it isn’t.
• Progress tracking: optional history so users can measure improvement over time (without making sharing or privacy feel risky).
• More languages and modes: support for different English prompt sets and possibly other languages in the future.

Built With

• nextjs
• typescripe