Kalenjin TTS โ€” Orpheus-3B fine-tune (v1)

First open-weights text-to-speech model for Kalenjin (kln), a Southern Nilotic language spoken by ~6.3M people in Kenya's Rift Valley. Fine-tuned from unsloth/orpheus-3b-0.1-ft via Unsloth on the AfriVoices-KE dataset.

Variant

Merged 16-bit (bf16) safetensors. Full standalone model (~6.6 GB) โ€” load directly with transformers. No separate base model needed at inference time.

Training summary

Base model unsloth/orpheus-3b-0.1-ft (3.3B Llama, Apache 2.0)
Data 66.9 hours of clean AfriVoices-KE Kalenjin scripted audio (42,106 clips)
Dialects Nandi 52.3% / Kipsigis 47.7% (anonymous multi-speaker)
Format 24 kHz mono PCM
LoRA rank 64 (rank-stabilized = false)
Target modules q/k/v/o + gate/up/down
Optimizer adamw_8bit, lr 2e-4 linear decay, bf16
Batch / grad accum 1 / 4 (effective batch 4)
Epochs 1.0
Final loss 4.42
Hardware NVIDIA H100 80GB (single GPU, ~2h)

Inference (with SNAC vocoder)

Orpheus emits audio-codec tokens that are decoded by hubertsiuzdak/snac_24khz. See the canonical inference recipe in Unsloth's Orpheus notebook.

Data attribution

Trained on AfriVoices-KE (Wanzare et al. 2026, CC BY 4.0) โ€” a Gates-Foundation- funded collaboration between Maseno University, USIU, Dedan Kimathi, Kabarak, JOOUST, KU Leuven, and Tech Innovators Network.

@misc{wanzare2026afrivoiceske,
  title={AfriVoices-KE: A Multilingual Speech Dataset for Kenyan Languages},
  author={Wanzare et al.},
  year={2026},
  eprint={2604.08448},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}

Limitations

  • Two dialect clusters only (Nandi, Kipsigis). Other Kalenjin varieties (Tugen, Marakwet, Sabaot, Sebei) are not represented.
  • Demographic skew: M 51.8% / F 35.1%, 71% tertiary-educated, age 18โ€“29 dominant. Model output reflects this distribution.
  • No held-out hidden-test evaluation โ€” dataset authors retain that 5% slice.
  • v1 release. No quality benchmarks vs. native speakers yet.

License

Inherits Apache 2.0 from the base model. Dataset is CC BY 4.0 โ€” attribution above. Derivative use must comply with AfriVoices-KE's prohibition on surveillance, discrimination, or exploitation.

Downloads last month
14
Safetensors
Model size
3B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Tonykip/kalenjin-tts-orpheus-v1

Paper for Tonykip/kalenjin-tts-orpheus-v1