---
title: Kinyarwanda ASR (from scratch)
emoji: 🎙️
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 6.18.0
app_file: app.py
pinned: false
license: cc-by-4.0
---

# Kinyarwanda ASR — Conformer-CTC, from scratch

Demo for the [LanguageBuild](https://github.com/Nsarob/LanguageBuild) project.
A ~20M-parameter Conformer encoder + CTC head, trained end-to-end from scratch on
Kinyarwanda speech — no pretrained weights. Greedy decoding.

**Honest scope:** trained on one in-domain corpus (image-description speech, single 2025
campaign). Real-mic, spontaneous, or health-domain speech is out-of-distribution and may
transcribe poorly. This is a research PoC, not a product.

## How the checkpoint is loaded

`app.py` loads the weights file `conformer-ctc-15-0.240.ckpt` from **the same folder as
app.py** — locally and inside the Space alike. No separate model repo is needed. (Optional:
set the env var `CKPT_REPO` to a HF *model* repo id to download from there instead.)


## 2. Deploy to the Space (browser, no CLI needed)

Space exists at  **huggingface.co/spaces/nsarob/kinyasr2026**.