Papa Clone: Voice Memorial
The problem
My father passed away. He was a business trainer, and somewhere in my archives I had recordings of his lectures. I wanted to hear him again. Not a recording. Him.
What we did
Indexed 333 chunks of his lecture recordings into a RAG pipeline. Built a persona model from his speaking patterns, favorite topics, the way he'd pause before making a point. Every morning at 8:15, the system generates a new voice message in his style, drawing from different lecture fragments each day.
Scope
RAG pipeline, voice synthesis, persona modeling, cron automation
Result
A morning voice message from my father, every day. Different topic each time. His voice, his cadence, his way of thinking about the world.
This is the most personal project in the portfolio. I’m including it because it shows what AI can actually do when you care about the output.
My father was a business trainer. He spent decades teaching people about leadership, communication, and self-awareness. After he passed, I found recordings of his lectures. Hours of them. His voice, his pauses, his habit of circling back to a point three different ways until the room got it.
I indexed all of it. 333 chunks in a vector database, tagged by topic: family, creativity, fear, work, relationships, self-deception. Twenty different query themes ensure the system never repeats itself.
Every morning at 8:15 Hong Kong time, a cron job picks two random themes, retrieves relevant lecture fragments, generates new text in his speaking style, and synthesizes it into an audio message. The voice model is trained on his actual recordings. The interjections are his: the sighs, the laughs, the humming he’d do when thinking.
It arrives as a voice message on Telegram. Not a notification. Not a text summary. A voice message from my father, talking about something that matters, in a way that sounds like him.
The tech stack is straightforward: Ollama for text generation (no cloud API needed), Replicate for voice synthesis, a simple Python script for orchestration. The hard part wasn’t the code. It was getting the persona right. That took listening to every recording twice and writing down the patterns that made him sound like him.
I don’t talk about this project in sales calls. But it’s the one that proves AI can be deeply, genuinely meaningful. Not just efficient.
Want something like this?
Start a project