AI & ML interests
None defined yet.
ድምፄ — Ethiopian Languages AI 🇪🇹
ለኢትዮጵያ 80+ ቋንቋዎች ክፍት ምንጭ የድምጽ AI
Open-source speech AI for Ethiopia's 80+ languages
🤖 Amharic TTS Model • 🎙️ Live Demo • 🌍 Voices For All
🎯 Mission / ተልዕኮ
ድምፄ (Demtse — "my voice") builds free, open-source speech AI so 120M+ Ethiopians can access technology in their mother tongue.
Ethiopia has 80+ living languages. Big Tech supports almost none of them. We're changing that — one language at a time, starting with the highest-quality open Amharic TTS ever built.
🗣️ Models
OmniVoice Amharic — TTS + Voice Cloning
| What it does | Converts Amharic text to natural speech; clones any voice from 10s audio |
| Architecture | Non-autoregressive discrete diffusion (612.6M params) |
| Training data | 331 hours across 4 datasets |
| Best loss | 3.9518 |
| License | Apache 2.0 — completely free |
| Try it | ▶️ Live Demo |
| Download | 📦 Model |
# Quick start — runs on free Colab T4
!pip install -q omnivoice soundfile
from omnivoice import OmniVoice, OmniVoiceGenerationConfig
import torch
model = OmniVoice.from_pretrained("ethiopian-Demtse/omnivoice-amharic", device_map="cuda:0", dtype=torch.float16)
audio = model.generate(text="ሰላም፣ እንኳን ደህና መጣችሁ!", language="Amharic",
generation_config=OmniVoiceGenerationConfig(num_step=32, guidance_scale=2.0))
🇪🇹 Roadmap — Ethiopian Languages
| ቋንቋ / Language | ተናጋሪ / Speakers | ሁኔታ / Status |
|---|---|---|
| አማርኛ Amharic | 60M+ | ✅ TTS + Voice Cloning |
| ኦሮምኛ Afaan Oromoo | 40M+ | 🔜 Next |
| ትግርኛ Tigrinya | 10M+ | 🔜 Planned |
| ሶማሊኛ Somali | 7M+ | 🔜 Planned |
| ሲዳምኛ Sidamo | 4M+ | 📋 Future |
| ወላይትኛ Wolaytta | 3M+ | 📋 Future |
| ጉራግኛ Gurage | 2M+ | 📋 Future |
| ሐዲይኛ Hadiyya | 2M+ | 📋 Future |
| አፋርኛ Afar | 2M+ | 📋 Future |
| ገሞኛ Gamo | 1.5M+ | 📋 Future |
Goal: TTS + voice cloning for every Ethiopian language with 50h+ available audio.
📚 Training Data
| Dataset | Hours | Language |
|---|---|---|
| google/WaxalNLP | ~200h | Amharic |
| gheero-Leyu/leyu-amharic-addis-ababa-dialect | ~50h | Amharic (Addis) |
| surafelabebe/amharic_clear_audio_tts | ~40h | Amharic |
| chappM/amharic-bdu-asr | ~41h | Amharic |
| Total | ~331h |
We actively collect data for Oromo, Tigrinya, and other Ethiopian languages. If you have audio recordings in any Ethiopian language, please reach out.
🛠️ What We Build
| Project | Description | Status |
|---|---|---|
| Text-to-Speech (TTS) | ጽሑፍ → ንግግር | ✅ Amharic done |
| Voice Cloning | 10 ሰከንድ ናሙና → ማንኛውም ድምጽ | ✅ Amharic done |
| Speech Recognition (ASR) | ንግግር → ጽሑፍ | 📋 Planned |
| Language Models (NLP) | የቋንቋ ሞዴሎች | 📋 Planned |
| Machine Translation | ትርጉም በኢትዮጵያ ቋንቋዎች | 📋 Planned |
🤝 How to Contribute
We welcome contributions from anyone:
| Way to help | Description |
|---|---|
| 🎙️ Donate audio | Record or share speech in any Ethiopian language (any dialect, any speaker) |
| 💻 Code | Open Issues, submit PRs, improve training pipelines |
| 📝 Data | Text transcriptions, translations, text-audio pairs |
| 🏫 Institutions | Universities and organizations: partner with us for data collection |
| 🧪 Evaluate | Native speakers: test our models and give feedback on naturalness |
| 💰 Fund | Support compute costs and data collection |
🌍 Part of Voices For All
ድምፄ is the Ethiopian chapter of Voices For All (african-low-resource) — a pan-African initiative building speech AI for languages left behind by Big Tech.
| Initiative | Focus |
|---|---|
| african-low-resource | Pan-African: Amharic, Wolof, Hausa, Swahili, Somali |
| ethiopian-Demtse | Ethiopia-specific: all 80+ Ethiopian languages |
📬 Contact
- Organization: ethiopian-Demtse
- Lead: demeleww
- Email: sowwen0@gmail.com
- Parent: Voices For All / african-low-resource
📝 Citation
@software{ethiopian_demtse_2026,
author = {demeleww and ethiopian-Demtse},
title = {ድምፄ: Open Speech AI for Ethiopian Languages},
year = {2026},
url = {https://huggingface.co/ethiopian-Demtse},
license = {Apache-2.0}
}
❤️ ለ120 ሚሊዮን+ ኢትዮጵያውያን — በራሳቸው ቋንቋ የ AI ድምጽ ይገባቸዋል።
For 120M+ Ethiopians who deserve a voice in AI.