Resurrecting Kekkonen: Training Qwen3-TTS Custom Voice

Introduction

Ever wanted to hear a historical figure speak in modern contexts? Yeah, me too. So I decided to train Qwen3-TTS to sound like Urho Kekkonen, Finland's long-serving president, using his iconic 1975 New Year's speech as training data.

The catch? I did this on a MacBook M4 with 24GB RAM. No fancy A100s. No cloud GPU clusters. Just me, my Mac, some ambition, and questionable decision-making skills.

The Setup: Teaching Qwen3-TTS About macOS

First problem: Qwen3-TTS didn't support macOS Metal Performance Shaders (MPS) out of the box. Since I wasn't about to rent a GPU farm, I rolled up my sleeves and added MPS support myself. (Yes, it is not merged yet. hurry up Qwen team.)

The Training Data: A Presidential Speech Meets DIY Audio Engineering

I didn't have access to massive audio datasets or professional recording studios. So I did what anyone with determination and questionable free time would do: I manually recorded 5-10 second audio clips, transcribed them, and used those as training data. I also snapped 10 second Kekkonen speech to be used as voice clone.

It was tedious. It was repetitive. It was exactly how many of history's great achievements began (hah).

The Results: Kekkonen Lives Again (Sort Of)

The good news: you can absolutely hear Kekkonen's voice. The characteristic cadence, the Finnish pronunciation patterns, the gravitas—it's all there.

The bad news: Due to the minimalist training dataset (GPU constraints, not laziness), the Finnish articulation isn't perfect. Some pronunciations are rough around the edges. But hey, that's what happens when you try to resurrect presidential authority with consumer hardware and prayer.

Hear It Yourself

Here are two example outputs from the trained model:

Example 1:

Example 2:

Example 3:

Notice the unmistakable Kekkonen cadence and vocal character. The training worked. The GPU jealousy was worth it.

Key Takeaways

• MPS support matters: Bringing TTS training to macOS opens doors for ML work outside data centers
• You don't need massive training sets to get recognizable results, but you do need some
• 24GB RAM on M4 is the real MVP (though it does complain about it)
• Fine-tuning vs. full training: We're not reinventing TTS here, just adapting it to a voice

What's Next?

With proper GPU access and a complete training dataset, you could get near-perfect Finnish pronunciation and smoother voice synthesis. For now, though, Kekkonen's digital resurrection is serving its purpose: proof that you don't need a Fortune 500 budget to do cool ML experiments.

Next I probably dig little bit deeper to improve model for more natural Finnish speech. Stay tuned (and be careful out there).