This is a research proof of concept, not a production-quality speech model. It was trained to explore whether the available Kalenjin speech data can support a usable first public Kalenjin TTS demo.

Live demo:

https://huggingface.co/spaces/mutaician/swahili-kalenjin-translator

Project code and attempt log:

https://github.com/mutaician/Kalenjin-TTS-V1

Model Details

Base architecture: F5-TTS v1 Base
Fine-tuning data: scripted Kalenjin subset from Anv-ke/Kalenjin
Training subset: speakers with at least about 1 hour of scripted audio
Approximate F5-prepared training set: 26,864 samples
Exported checkpoint: non-EMA model_state_dict
Exported checkpoint update: 19,980
Tokenizer/vocab: custom F5-TTS vocab from the fine-tuning run
Vocoder used for inference: Vocos 24 kHz

Intended Use

This model is intended for:

Kalenjin speech synthesis research
proof-of-concept demos
low-resource language experimentation
community feedback and dataset/model iteration

It is not intended for high-stakes, production, identity-sensitive, or misleading voice generation use.

Limitations

Pronunciation and prosody can be wrong.
Short text can sound less natural than longer sentences.
The model may skip words or speak too quickly for some prompts.
Speaker consistency depends on the reference audio used at inference time.
The training data was released for ASR and related research, not specifically curated as a polished TTS dataset.

Files

model_last.safetensors: exported non-EMA model weights
vocab.txt: custom vocab used by the fine-tuned model
export_info.json: export metadata from the Modal training/export pipeline

Reference audio is not included in this repo. F5-TTS inference requires a reference audio clip and exact reference transcript. The live demo uses one curated internal reference clip.

Local Inference

Install F5-TTS:

pip install f5-tts soundfile

Example:

bash

f5-tts_infer-cli \
  --model F5TTS_v1_Base \
  --ckpt_file model_last.safetensors \
  --vocab_file vocab.txt \
  --ref_audio path/to/reference.wav \
  --ref_text "Exact text spoken in the reference audio." \
  --gen_text "Kalenjin text to synthesize." \
  --output_dir outputs \
  --output_file kalenjin_tts.wav \
  --nfe_step 32 \
  --cfg_strength 2.0 \
  --speed 0.8

The local wrapper used during development is available in the project code:

6_local_f5tts_inference.py

Dataset and Acknowledgements

This experiment depends on data work from Africa Next Voices and the Anv-ke/Kalenjin dataset. Thanks to the Africa Next Voices team and the people who contributed to the Kalenjin speech data effort.

Disclaimer

This model is experimental. Generated speech may be inaccurate, unclear, or unnatural. Please treat it as a research artifact and community demo.

mutaician/kalenjin-f5tts-v1

Kalenjin F5-TTS