
Base Model	facebook/mms-tts-kik
Architecture	VITS (Variational Inference with adversarial learning for end-to-end Text-to-Speech)
Language	Kikuyu (kik) — Bantu, tonal language
Parameters	~36M
Training Data	48,500 audio samples from ANV + WAXAL datasets
Training	GAN-based adversarial fine-tuning on A100 80GB GPU
Developer	C-elo Labs

Training Data

This model was fine-tuned on real human Kikuyu speech recordings from two sources:

Dataset	Samples	Description
ANV (African Next Voices)	~46,900	Scripted Kikuyu speech across 5 dialects
WAXAL (Google)	~1,600	Studio-quality TTS recordings

Usage

python

from transformers import VitsModel, AutoTokenizer
import torch
import scipy.io.wavfile as wavfile

model = VitsModel.from_pretrained("gateremark/kikuyu-tts-v1")
tokenizer = AutoTokenizer.from_pretrained("gateremark/kikuyu-tts-v1")

text = "ũhoro waku"
inputs = tokenizer(text=text, return_tensors="pt")

with torch.no_grad():
    output = model(**inputs)

waveform = output.waveform[0].numpy()
wavfile.write("output.wav", model.config.sampling_rate, waveform)

Training Details

Method: Fine-tuned using the ylacombe VITS training recipe with GAN-based adversarial loss
Epochs: ~70,000 steps
Optimizer: AdamW
Loss: Mel reconstruction + KL divergence + discriminator adversarial loss
Hardware: NVIDIA A100 80GB (Modal)

Limitations

Optimized for standard Kikuyu text; may struggle with code-switched (Kikuyu-English) input
Tonal accuracy depends on correct diacritical marks in input text (e.g., ũ vs u)
Single-speaker output

About C-elo Labs

C-elo Labs is an AI research organization pioneering high-fidelity language models for underserved African languages. We build text and voice AI systems for Kikuyu, Kamba, and other African languages.

Citation

If you use this model, please cite:

bibtex

@misc{gatere2026kikuyutts,
  title={Fine-tuned Kikuyu TTS based on MMS-TTS},
  author={Mark Gatere},
  year={2026},
  publisher={C-elo Labs},
  url={https://huggingface.co/gateremark/kikuyu-tts-v1}
}