mutaician/Kalenjin-TTS-V1

# Kalenjin-TTS-V1 An experimental first attempt at building a Kalenjin text-to-speech model. This repository is not a polished training framework. It is a record of the steps, mistakes, fixes, and findings from a fast engineering attempt to answer one question: > Can the available Kalenjin speech data be pushed far enough to create a > usable public TTS proof of concept? The answer from this attempt is: yes, but with clear limitations. Before this Modal/F5-TTS attempt, I had made smaller personal attempts to build Kalenjin TTS locally. Those attempts did not get far enough to be useful, unlike my earlier translation work. This repo starts from the point where I accepted that Kalenjin TTS needed a larger dataset, better tooling, and cloud GPU training. Live demo: huggingface.co Model: mutaician/kalenjin-f5tts-v1 ## Status This project produced: - a public Kalenjin TTS proof-of-concept model - a Hugging Face Space that supports translation plus optional Kalenjin speech - Modal scripts for dataset caching, preparation, training, export, and testing - notes on what worked, what failed, and where future attempts should improve This project did not produce: - a production-quality TTS model - a fully evaluated benchmark - a clean reusable training library - a guarantee of correct pronunciation or prosody ## Budget And Hardware This attempt used Modal cloud GPUs. A practical budget for reproducing this style of experiment is about: ```text $100 in Modal credits ``` That budget covered failed runs, dataset preparation, one top-4-speaker fine-tune, one broader top-1-hour-speaker continuation run, export tests, and basic deployment experiments. A careful reproduction could cost less; systematic experimentation across many configurations would cost more. The main training runs used A100 80GB GPUs. The public inference demo is designed to use a cheaper T4 GPU and scale to zero when i …