SpeakFast | Devpost

Inspiration

Coming from immigrant families, we’ve seen firsthand how difficult it can be to fully express yourself in English as a second language. We’ve watched our parents and many others struggle to reach their full potential because of communication barriers.

We also noticed that most existing language-learning tools don’t feel realistic. They often rely on predictable prompts or multiple-choice answers, which can make practice feel artificial. We wanted to change that by creating a tool that simulates real, spontaneous conversations, helping people build confidence in how they actually speak, not just how they study.

What it does

SpeakFast randomly calls users and engages them in surprise AI-driven conversations, ranging from real-world situations to imaginative, fun, fantasy scenarios. By simulating unpredictable speaking contexts, it helps users practice English naturally, build confidence, and become more comfortable expressing themselves on the spot.

How we built it

We built SpeakFast in 24 hours with a four-person team — two focusing on the frontend and two on the backend. On the backend, we implemented a multi-API pipeline that connects speech, language understanding, and synthesis in real time:

ElevenLabs API for converting AI responses into natural text-to-speech audio

Google Cloud Speech-to-Text for accurately transcribing user speech inputs

Gemini API for generating context-aware conversation responses and grammar feedback

The frontend was developed in React Native and deployed with Expo Go, providing a consistent, cross-platform user experience for both iOS and Android. We also integrated audio recording, live playback, and asynchronous conversation handling to simulate a realistic AI phone call experience.

Challenges we ran into

One of the toughest challenges was designing AI prompts that could guide users through conversations while still leaving room for open-ended learning and practice. Balancing structure and flexibility took iteration.

Another major challenge was juggling three heavily interconnected APIs. ElevenLabs, Google Cloud, and Gemini all needed to work seamlessly together, and small issues in one service often cascaded into larger problems. Debugging and syncing them was time-consuming but rewarding.

Accomplishments that we're proud of

We’re proud that we were able to complete our MVP within the hackathon and delivered a fully functional app that doesn’t just look polished but actually works end-to-end. Watching it handle full conversations in real time was a huge milestone for us.

What we learned

We learned how important it is to design prompts that feel conversational rather than robotic, and how much of AI’s effectiveness depends on the framing of interactions. We also gained experience in coordinating multiple cloud services, managing dependencies, and building a system resilient enough to run in real time.

From a teamwork perspective, we learned the value of rapid task division, quick feedback loops, and leaning on each other’s strengths to keep the project moving.

What's next for SpeakFast

The number one priority for SpeakFast is enhancing the interface to support multiple languages, allowing users to navigate and interact seamlessly in their native language. While we focused on English for this hackathon, our next goal is to expand SpeakFast into a global language learning platform, introducing AI-driven practice for other languages such as Spanish, Korean, Mandarin, and French.

Next, we want to fine-tune the Gemini models to provide better guidance and more natural conversation flow. We’re also exploring ways to give learners more effective feedback at the end of each session, including better vocabulary suggestions with finer-tuned models.

Ultimately, we hope SpeakFast can empower more people to practice speaking English, one surprise call at a time.

The two github repos are for the frontend (hosted on react native) and backend (hosted on github railway)