README.md

metadata

title: Say It Back in Kalenjin
emoji: 🗣️
colorFrom: yellow
colorTo: red
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false
short_description: Say a word in English, hear it in Kalenjin
tags:
  - build-small-hackathon
  - backyard-ai
  - off-brand
  - tiny-titan
  - best-demo
  - bonus-quest-champion
  - modal
  - speech
  - low-resource-languages
  - kalenjin
  - track:backyard
  - sponsor:modal
  - achievement:welltuned
  - achievement:offbrand

🗣️ Say It Back in Kalenjin

Say a word in English. Hear it in Kalenjin, in a real Kalenjin voice.

Kalenjin is slipping away in my family. My wife thinks in English and wants to learn it; this lets her say a word and hear how to say it back in Kalenjin.

How it works

Speak (or type) English → Whisper transcribes it → an English→Swahili→Kalenjin cascade translates it → Cheps (a Kalenjin TTS I fine-tuned) speaks it. Common phrases are pre-rendered so they play instantly.

Build Small (Hugging Face × Gradio) — what we're going for

• Track: 🏡 Backyard AI — built for a real person (my wife, learning Kalenjin).
• 💚 Modal — Best Use of Modal: the translation cascade and the Cheps voice both run on Modal at runtime; this Space orchestrates them.
• 🎨 Off Brand: a fully custom UI (editorial serif, terracotta on off-white, the Kalenjin word as the centerpiece) — well past the default Gradio look.
• 🐣 Tiny Titan: every model is ≤4B — Whisper 74M (ASR), NLLB 600M ×2 in the cascade (translation), Orpheus 3B + LoRA (the Cheps voice). Small weights, real impact for a language with almost no tooling.
• 🏆 Best Demo: app + demo video + social post.
• 👑 Bonus Quest Champion: stacking the most bonus criteria above.
• Demo video: https://www.instagram.com/reel/DZn70fHyNa-/
• Social post: https://www.instagram.com/reel/DZn70fHyNa-/

Models run on Modal; this Gradio Space ties them together.