Hope226/amharic_tts

🎙️ Amharic Text-to-Speech — facebook/mms-tts-amh

A fully free, offline Amharic TTS app using Meta's MMS model, running locally via Gradio.
No API keys. No internet connection required after the model is downloaded.

---

📁 File Structure

amharic_tts/
├── app.py            ← main application (single file)
├── requirements.txt  ← Python dependencies
└── README.md         ← this file

---

🖥️ Local Setup (Step-by-Step)

Prerequisites

• Python 3.9 – 3.12 (3.10 recommended)
• ~2 GB free disk space (model cache + packages)
• No GPU required

---

Step 1 — Clone / download the files

Place app.py and requirements.txt in a folder, e.g. amharic_tts/.

mkdir amharic_tts && cd amharic_tts
# copy app.py and requirements.txt here

---

Step 2 — Create a virtual environment

# Create
python -m venv venv

# Activate — macOS / Linux
source venv/bin/activate

# Activate — Windows (PowerShell)
venv\Scripts\Activate.ps1

---

Step 3 — Install dependencies

pip install --upgrade pip
pip install -r requirements.txt

First install takes 3–8 min because it downloads the CPU PyTorch wheel (~220 MB).

---

Step 4 — Run the app

python app.py

You should see:

Pre-loading model at startup …
⏳  Loading facebook/mms-tts-amh — this may take a minute on first run …
✅  Model ready  (sampling_rate=16000 Hz)
Running on local URL:  http://0.0.0.0:7860

First run downloads ~330 MB from Hugging Face Hub and caches it in
~/.cache/huggingface/hub/.
Subsequent runs start in seconds.

Open your browser at http://localhost:7860 and start typing!

---

Troubleshooting

Symptom	Fix
ModuleNotFoundError: uroman	pip install uroman
No module named 'transformers'	pip install transformers>=4.40.0
Slow first generation	Normal — model loads on first request if pre-load failed
Port 7860 already in use	Change server_port=7861 in app.py
OMP: Error #15 on macOS	export KMP_DUPLICATE_LIB_OK=TRUE before running

---

🚀 Free Deployment on Hugging Face Spaces

Hugging Face Spaces gives you a completely free public URL with 2 vCPUs and 16 GB RAM — plenty for CPU inference.

Step 1 — Create a Space

1. Go to huggingface.co/spaces and sign in.
2. Click New Space.
3. Set:
   • Space name: amharic-tts (or anything you like)
   • SDK: Gradio
   • Hardware: CPU Basic (free tier)
   • Visibility: Public or Private

Step 2 — Upload your files

Upload both files via the web UI "Files" tab, or use the Git CLI:

# Install HF CLI if needed
pip install huggingface_hub

# Login
huggingface-cli login

# Clone your empty space
git clone https://huggingface.co/spaces/YOUR_USERNAME/amharic-tts
cd amharic-tts

# Copy your files in
cp /path/to/app.py .
cp /path/to/requirements.txt .

# Push
git add .
git commit -m "Initial Amharic TTS app"
git push

Step 3 — Wait for the build

Spaces automatically installs requirements.txt and runs app.py.
Watch the Logs tab — build takes 3 min.
First synthesis downloads the model (330 MB) into the Space's cache.

Step 4 — Share!

Your app is now live at:
https://huggingface.co/spaces/YOUR_USERNAME/amharic-tts

---

⚙️ How uroman Works

The MMS-TTS model was trained on romanised (Latin-script) phonetic text, not raw Ethiopic script. The uroman library converts:

ሰላም! → sälam!

transformers >= 4.36 ships built-in support: the VitsTokenizer detects that facebook/mms-tts-amh needs uroman and calls it automatically during tokenizer(text, return_tensors="pt").

If uroman is not installed you'll get an ImportError with a clear message.

---

📜 Licence

• Model: CC-BY-NC 4.0 (non-commercial)
• This app code: MIT