Voices For All — Open Voice AI for Low-Resource Languages

Mission: Every language deserves a voice. We build open, state-of-the-art speech AI for the world's 7,000+ languages — starting with those left behind by Big Tech.

🌍 Why This Matters

Over 90% of the world's languages have zero access to modern text-to-speech (TTS), speech recognition (ASR), or voice assistants. The biggest AI labs focus on English, Mandarin, Spanish, and a handful of other high-resource languages. Meanwhile:

• Amharic (~60M speakers) has only a handful of open TTS models
• Wolof (~12M speakers) has virtually no open TTS with voice cloning
• Hausa (~90M speakers) has critically underserved speech tools
• Swahili (~200M speakers) has basic TTS but no open voice cloning
• Somali, Zulu, Igbo, Yoruba — all left behind

This is not a technical limitation. It is a funding and prioritization gap. We exist to close it.

🏗️ Our Approach

🎯 Proof of Concept: Amharic OmniVoice

Our first model is a fully functional, high-quality Amharic TTS + Voice Cloning system:

• Model: african-low-resource/omnivoice-amharic
• Architecture: Non-autoregressive discrete diffusion (OmniVoice, 612M params)
• Training Data: 81K samples, 331 hours across 4 curated Amharic datasets
• Capabilities: Text-to-speech, zero-shot voice cloning, multi-speaker synthesis
• Best Loss: 3.9518 (state-of-the-art for this architecture on Amharic)
• License: Apache 2.0
• Demo: Live Gradio Space

Why This Proves We Can Scale

• Data pipeline is reusable: Tokenization, cleaning, and training scripts work for any language with ~100h of audio
• Model is lightweight: 612M params runs on a free T4 Colab — no API keys, no cloud bills
• Voice cloning works: A speaker can clone their voice with 10 seconds of audio
• Community validated: Native Amharic speakers confirm natural prosody

🗺️ Roadmap: 2026–2028

🤝 Partners & Funders

We are actively seeking partnerships with:

• Grant bodies: Mozilla Common Voice, Lacuna Fund, Gates Foundation, IDRC
• Research networks: Masakhane, AI4D Africa, GalsenAI, Google Research Africa
• Local institutions: Ethiopian AI institutes, East African universities
• Industry: Telecoms, banking, agriculture tech building voice interfaces

📊 Impact Metrics We Track

📬 Contact

• Lead: demeleww
• Organization: Voices For All
• Hugging Face: huggingface.co/african-low-resource

"Technology should not be a privilege of the languages with the most data. It should be a right for every community that speaks."