Namuai-org/namu-tts-amharic-tts-internship

# Amharic Text-to-Speech — Namu Internship Project **Goal:** every intern fine-tunes a text-to-speech model that speaks Amharic (አማርኛ), listens to the result, and measures honestly whether it got better. **Team:** 6 interns, one model each · **Duration:** 2 weeks (10 working days) · **Level:** beginner (no prior speech-ML experience assumed) This is **not** an attempt at state of the art, and it never will be. It is an introduction with three aims: see how fine-tuning actually works, hear the result of your own work, and experiment with a real TTS model. --- ## Read in this order | # | Document | What it gives you | |---|---|---| | 0 | Project brief | Why this project, what "done" means, what is deliberately not the goal | | 1 | Background | Amharic + TTS primer, what to read on which day | | 2 | Datasets | Every candidate corpus, measured, and which one was pre-picked for you | | 3 | Team & tasks | Six individual owners, three cross-cutting duties, review buddies | | 4 | Timeline | Day 0 pre-bake plus 10 days, with the gate at the end of each | | 5 | Evaluation | Noise floor, ASR floor, CER harness, MOS-lite blind A/B | | 6 | Setup | Colab-only environment, exact commands | | 7 | **Gotchas** | **The traps that will eat your week. Read before you touch anything.** | | 8 | Assessment rubric | How the work is assessed | | 9 | Experiments | The six axes: hypothesis, method, expected result, cost | --- ## The 60-second version We fine-tune `facebook/mms-tts-amh` — an 83M-parameter VITS model that already speaks Amharic, badly — on a cleaned single-speaker Amharic corpus, using `ylacombe/finetune-hf-vits`. One run is about 20–25 minutes on a free-tier T4, and the upstream repo reports usable results from as few as 80–150 samples. That is why you get many cycles instead of one. The hard part is **not** the training. It is the data and the text pipeline. Three facts we measured while writing this: 1. **`mms-tts-amh` cannot read Amharic script.** Its vocabulary i …