Kasanoma – Offline TTS Models for languages of Africa

Kasanoma is an open-source initiative dedicated to building and sharing offline Text-to-Speech (TTS) models for languages of Africa.
With Kasanoma models, you can generate natural-sounding voices entirely on your local device — no internet connection required.

All models in this project are built using Piper, a fast, lightweight neural TTS system that runs efficiently on laptops, desktops, Raspberry Pi, and other low-resource devices.

You can also try out the models live online at our demo site. The models are trained for a limited amount of time due to GPU resource availability but they can be finetuned for better results using the training checkpoints which are also published.

---

Available Models

We currently provide offline TTS models for the following languages:

Language	Main Countries	Native Speakers	Total Speakers
Twi (Akan)	Ghana	~9–10 million	~17 million
Chichewa	Malawi, Zambia, Mozambique, Zimbabwe	~12 million	~18 million
Makhuwa	Mozambique	~7–8 million	~8 million

You can download the models from releases.

More African languages will be added soon.
Languages in the pipeline: Kikuyu, Tshiluba, Amharic, Vai.

---

How to Use the Models Offline

All Kasanoma models are compatible with Piper.
Here’s how to run them without internet access:

1. Install Piper
   You can install it via pip:

   pip install piper-tts

   Or download a standalone binary from the Piper releases page for your platform (Linux, macOS, Windows).

2. Download a Kasanoma Model
   Get the .onnx model file and its matching .json config file for your language.

3. Run TTS

   piper --model kasanoma-twi-medium.onnx --output audio.wav <<EOF
   Me pɛ sɛ me kɔ sukuu no mu.
   EOF

   This will generate audio.wav with natural Twi speech — fully offline.

---

Contributing

We welcome contributions in the form of:

• Training new models for other languages of Africa or finetuning the released models. All the knowledge from training models is open source and will be shared freely to support you.
• Audio quality testing and pronunciation feedback.
• Add scripts for using the models offline

If you have any questions or would like to contribute, our email is kasanoma@kasanoma.org.

---

License

All Kasanoma models are released under open-source licenses.