An open-source neural text-to-speech model for Ng'akarimojong (kdj) — an Eastern Nilotic language spoken by approximately 370,000 people in the Karamoja sub-region of northeastern Uganda.

One of the first open-weights TTS systems for this language.

🔊 Samples

Karamojong	English	Model	Audio
"Abaraka noi ejok akim"	"Thank you very much, it is good"	best_model	`sample_1.wav`
"Ijo noi Anu ejok noi"	"Yes indeed, Anu is fine indeed"	best_model	`sample_2.wav`
"Abunit ayong Najie anerai acamit ayong akisyom ngiemuto ka ngitalio a Ngijie"	"I came to Najie because I want to learn about the history and culture of the Jie"	best_model	`samples/best_model__abunit_ayong_najie_anerai_acamit_ayong_akisyom_ngiemuto.wav`
"Abunit ayong Najie anerai acamit ayong akisyom ngiemuto ka ngitalio a Ngijie"	"I came to Najie because I want to learn about the history and culture of the Jie"	best_model_9918	`samples/best_model_9918__abunit_ayong_najie_anerai_acamit_ayong_akisyom_ngiemuto.wav`

🌍 Why this matters

Of the world's ~7,000 languages, fewer than 100 have meaningful speech technology. Karamojong — despite hundreds of thousands of speakers — has been almost entirely absent from voice assistants, accessibility tools, and language-learning apps. This model is a step toward changing that.

🚀 Quick start

CLI

bash

# Install with pinned versions to avoid dependency conflicts
pip install \
    "coqui-tts==0.27.5" \
    "transformers==4.57.6" \
    "tokenizers>=0.22,<0.23" \
    "numpy>=2.0,<2.3" \
    "numba>=0.61,<0.62" \
    "librosa>=0.11,<0.12" \
    "huggingface_hub>=0.34,<0.40"

# Pull model + config from the Hub
python -c "from huggingface_hub import snapshot_download; \
    snapshot_download('speedykom-group/karamojong-tts-vits-beta', local_dir='./karamojong-tts')"

# Synthesize
tts --text "Ejok noi! Iyalama ngesi?" \
    --model_path ./karamojong-tts/best_model.pth \
    --config_path ./karamojong-tts/config.json \
    --out_path output.wav

Python

python

from huggingface_hub import snapshot_download
from TTS.utils.synthesizer import Synthesizer

repo_dir = snapshot_download("speedykom-group/karamojong-tts-vits-beta")

syn = Synthesizer(
    tts_checkpoint=f"{repo_dir}/best_model.pth",
    tts_config_path=f"{repo_dir}/config.json",
)

wav = syn.tts("Ejok noi! Iyalama ngesi?")
syn.save_wav(wav, "output.wav")

🏗️ Model


Architecture	VITS (end-to-end TTS)
Framework	Coqui TTS
Sample rate	22,050 Hz, mono WAV
Speakers	Single-speaker
Language	Ng'akarimojong (ISO 639-3: `kdj`)
Family	Eastern Nilotic → Ateker (Teso–Turkana)
License	CC BY-NC 4.0

🧪 Training

Trained on a single NVIDIA L4 GPU (24 GB) with Coqui TTS. Key design choices:

Source data — 4,622 clips (1–8 s, ~5–6 hours total) from speedykom-group/karamojong-speech-dataset.
Audio preprocessing — Stationary noise reduction via noisereduce applied to the source audio. Cleaner training signal translates directly to more natural output.
Loss balancing — mel_loss_alpha=40, feat_loss_alpha=2.5, gen_loss_alpha=1.5. The boosted feature-matching weight pushes outputs from "correct mel-spectrogram" toward "naturally human."
Schedule — Cosine annealing with warm restarts; balanced generator/discriminator learning rates.

Beta-15 is the latest checkpoint in an iterative training process and represents the current best perceptual quality.

Early training history (beta-0 → beta-13) on raw audio. Beta-15 uses a denoised dataset — values aren't directly comparable.

🤝 Dataset & ethics

Training data comes from the Karamojong Speech Dataset by Speedykom Group, derived from publicly available GRN recordings. This work is offered as a contribution to digital preservation and accessibility for an underrepresented language community. The non-commercial license reflects the spirit of the source material.

👥 Credits


Dataset	Speedykom Group
Model training	@dhiakassab
TTS framework	Coqui TTS
Hardware	NVIDIA L4 GPU (24 GB) ×1

📄 License

CC BY-NC 4.0 — free for research and non-commercial use with attribution to Speedykom Group. For commercial licensing, contact Speedykom Group.

Part of an effort to build speech technology for underserved African languages.