This is a submission for the Hacktoberfest Open-Source AI Challenge: Week 2 — Side Hustle.
What I Built
EnglishChunks.guru is a 12-minute daily speaking coach for English learners.
Many learners know the words but freeze when they have to speak. Native speakers don't build sentences word by word — they use chunks like “make a decision” or “that makes sense”. So every day the app gives you one chunk at your level (A1–C2) and walks you through it:
- Learn — hear it and see three real examples
- Speak (twice) — record yourself, get a 0–100 score and a tip on the exact sounds to fix
- Write — use it in three sentences of your own; they come back polished
- Read aloud — say your sentences and get today's score
Who would pay: adults who need English for work — interviews, meetings, moving abroad. Plans are $7–$11/month (1–5 practices a day). It's free while I test it.
Demo
👉 Try it: https://englishchunks.guru (sign up with Google or email)
Code
gusanchefullstack
/
englishchunks.guru
Learn real English chunks daily with self-hosted open-source pronunciation scoring
EnglishChunks.guru
A web app that teaches English chunks — the ready-made phrases native speakers use, like “that makes sense” or “get used to” — to learners from A1 to C2, with one daily session: hear it, say it twice, write it, read it aloud, and get pronunciation scores and friendly tips from open-source AI running on infrastructure we control.
Live: https://englishchunks.guru
Table of Contents
- Why EnglishChunks.guru?
- Screenshots
- How it works
- Installation
- Quick Start
- Configuration
- API Reference
- Project Structure
- Tests
- Deployment
- What I learned
- Roadmap
- Contributing
- License
- Credits
- Author
Why EnglishChunks.guru?
Most apps teach single words, but fluent speakers think in chunks: “by the way”, “make a decision”, “I tend to…”. Learners need to say those chunks out loud and know exactly which sounds to fix — and pronunciation feedback usually means sending your voice to a closed cloud API.
This project does the speech scoring with…
How I Built It
The open models do the core job — listening to your voice and coaching you:
| Job | Open model | Runs on |
|---|---|---|
| Find your speech | Silero VAD | my own server |
| What you said | NVIDIA Parakeet TDT 0.6B (int8 ONNX) | my own server |
| How each sound was pronounced | wav2vec2 phoneme model + forced alignment | my own server |
| Friendly tips & writing feedback | Gemma 4 26B (open weights) | Vercel AI Gateway |
| Checks every library chunk | gpt-oss-20b as an independent auditor | Vercel AI Gateway |
- The scoring service is Python/FastAPI on one DigitalOcean Droplet (2 vCPU, 4 GB). A score comes back in ~1.1 s on a typical recording.
- The web app is Next.js 16 on Vercel with Postgres on Neon; audio examples come from ElevenLabs and are cached on DigitalOcean Spaces.
- The AI never invents the teaching: pronunciation hints come from a curated table of ~40 English sounds, and the model only wraps them in a warm sentence.
Why Does Open Innovation Matter?
- Privacy: your voice never goes to a third-party API. Recordings are scored in memory on my server and deleted immediately.
- Cost: a cloud speech API charges per minute of audio. Self-hosted models cost a flat $28/month, which is what makes $7 plans possible.
- Control: I can see why a sound scored low (phoneme-level alignment) and calibrate it on a public, expert-rated dataset — it ranks the better pronunciation higher in 89% of pairs. That's impossible with a black-box API.
Prize Categories
- Best Use of Gemma — Gemma 4 writes every tip and grades every sentence.
- Best Use of DigitalOcean — the self-hosted speech scoring runs on a Droplet; audio and avatars live on Spaces + CDN.
- Best Use of ElevenLabs — every chunk, example and learner sentence is spoken by ElevenLabs voices (generated once, then cached).
Thanks to the Hacktoberfest team for creating these challenges! A lot of learnings from every week! 👏👏👏💪
Top comments (0)