Nunsi Shiaki
AI/ML EngineerNone
Nunsi Shiaki is a Computer Engineering student, Data Scientist, and AI/ML Engineer passionate about building practical AI solutions that create real-world impact, particularly in low-resource environments. Her work spans machine learning, natural language processing, computer vision, and applied AI research, with projects focused on education, agriculture, healthcare, and social impact across African contexts.
Beyond building AI systems, Nunsi is deeply committed to growing the technology community. She actively contributes through organizations such as DSN, IEEE, Zindi, and Microsoft student programs by organizing learning events, mentoring aspiring practitioners, and making AI and Python more accessible. She enjoys breaking down complex technical concepts into simple, relatable lessons that help beginners build confidence while encouraging experienced learners to think more deeply about applying AI and Python to solve real-world problems. Driven by growth, collaboration, and responsible innovation, she is passionate about using technology to empower people and communities.
Abstract
When I first looked at speech synthesis, I assumed it was completely off-limits, requiring massive compute, huge datasets, and deep specialized knowledge. Then, I decided to just try it anyway. In this talk, I will show you how I built a custom voice cloning system from scratch using Python and a free Google Colab T4 GPU. We’ll walk through my exact pipeline: from curating a custom Hugging Face dataset (https://huggingface.co/datasets/Nunsi/tts-data), to pulling in the 1-billion parameter CSM-1B model via Unsloth, and setting up rank-stabilized LoRA fine-tuning. More importantly, we will dive into the messy reality of working with audio data in Python, the stuff standard tutorials usually skip. I will share how I figured things out by breaking them, including: - Preventing Memory Crashes: How to manually calculate and truncate audio arrays to safe 20-second limits (480,000 samples at a 24kHz sampling rate) so you donot run out of RAM. - Surviving Environment Bugs: How to stop your Colab runtime from completely dying by forcefully uninstalling broken libraries like torchcodec and safely routing everything through a soundfile backend. - The Training Loop: What it actually looks like to train this model, track memory usage, and save your fine-tuned adapters locally or to Hugging Face. We will wrap up by running an interactive inference session, playing a side-by-side comparison of the original ground-truth voice versus the newly cloned text-to-speech result. This session is for students, language enthusiasts and Python developers who are curious about speech AI but may not have worked with audio before. If you know Python and have touched machine learning even once, you will be able to follow along completely, and you will leave knowing exactly how to start your own voice cloning project.