Urgent.News

What's breaking now, across thousands of outlets.

AI

Misspoke

This is a submission for the Hacktoberfest Weekend Challenge: Build for a Friend . What I Built I built MisSpoke , a voice-first AI language practice partner for a friend who wanted to get better at speaking without every conversation feeling like a grammar test. Most language-learning apps are great at vocabulary, streaks, and exercises, but speaking is a different problem. You can know what a…

This piece is a contribution for the Hacktoberfest Weekend Challenge: Build for a Friend. The application I developed is named MisSpoke, an AI language practice companion designed for an individual seeking to improve their speaking skills without every conversation feeling like a grammar test. While numerous language-learning apps excel in vocabulary, streaks, and exercises, they fall short when it comes to actual speaking.

Despite knowing a sentence's meaning, one might still freeze when it's time to utter it. Hence, MisSpoke is centered around conversation. Users select the language they are learning, their proficiency level, and the desired tutor personality, and then begin speaking. The tutor can exhibit various personas, including calm, casual, motivating, energetic, creative, or a few others, depending on what makes the practice less awkward.

The app also retains the conversation, translating phrases when needed, tracking errors and recurring mistakes, and keeping the session flowing without interrupting every few seconds with a red pen.

To provide a clearer picture, here's how the application operates:

1. To start, users must choose their native and target languages, select their proficiency level, and choose a tutor personality.

2. They can then initiate a voice conversation.

3. At any point, they can request translations or explanations if they get stuck.

4. The application keeps track of the conversation and learning signals throughout the session.

5. If voice input is impractical, users can switch to text chat.

The development of MisSpoke involved a combination of technologies:

- Next.js 15

- React 19

- TypeScript

- ElevenLabs Conversational AI for real-time voice conversations

- OpenRouter for the language-model layer

- Supabase for authentication and storage

- IndexedDB for local session state and learning memory

- Framer Motion + Tailwind CSS for the interface

The app includes distinct sections for chat, tutor-prompt generation, translation, UI translation, and session summaries. MisSpoke leverages ElevenLabs to manage the conversational voice layer, enabling users to speak instead of engaging in another typing exercise. The tutor's behavior adapts depending on the learner's chosen style.

Current personalities include cheerful, calm, motivating, intellectual, creative, casual, empathetic, and energetic. This isn't merely cosmetic; the personality influences how the tutor responds and how the practice session feels.

MisSpoke supports multiple learning languages, including Spanish, French, Japanese, German, Italian, Korean, Mandarin, Hindi, Tamil, Telugu, Malayalam, and Kannada. If users find themselves stuck during a conversation, they can translate a tutor response instead of abandoning the session and using another app. It maintains a record of several elements such as conversation history, recurring mistakes, vocabulary learned, corrections, confidence levels, strengths, weaknesses, and session summaries.

Some of this learning data is saved locally using IndexedDB, while Supabase handles account-level storage. This setup ensures the tutor doesn't behave as if it has amnesia each time the page is refreshed. By making the model layer replaceable, MisSpoke allows for experimentation with open-weight models through the same application rather than rebuilding the product around a single provider.

This approach is particularly beneficial for language learning, as conversations can grow remarkably personal. They cover topics like families, college, work, embarrassing mistakes, plans, opinions, and more that may arise while practicing. With the use of open models and local inference, the app aims to keep far more of this processing on the learner's device.

Furthermore, it can eventually support locally running models, offline inference, fine-tuned language tutors, region-specific models, different models for different target languages, and custom correction behavior.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in AI

FocusBuddy: A Tiny AI Companion That Helps My Friend Start

What I Built I built FocusBuddy for a friend who wants to study and work but finds it hard to start when scrolling is always one tap away.

  • FocusBuddy aids friend in starting tasks by offering two-minute commitment.
  • Tracks user distraction with 60-second scroll breaks and phone usage pauses.
  • Uses open-weight AI model for optional nudge, stored locally without external transmission.

More from Saturday 3 October →