Today I had my ego gently dismantled by an AI. And honestly? I'm here for it.
I'm building an audio platform to explain health conditions to patients in their mother tongue. The goal is simple: make medical knowledge accessible to anyone, in the language they think in.
What I didn't expect was the humbling lesson I'd learn along the way.
The Setup
I needed Tamil voiceovers to explain conditions like diabetes, hypertension, and heart disease to patients. Being a Tamil speaker myself (albeit one whose pronunciation makes my relatives wince after 20 years in the UK), I assumed I'd either use my own voice or find native Tamil voice actors.
Enter ElevenLabs and their multilingual AI models.
The Plot Twist
After testing dozens of voices, including native Tamil speakers, I made a discovery that stopped me in my tracks:
The best Tamil voices weren't Tamil at all.
- Carmelo — an Italian e-learning professional
- Bradley — an American narrator
- Shelly — a British voice artist
- Nikalia — an American with a "bright and friendly" style
These voices, originally trained on English and European languages, produced better Tamil than the native options. Clearer pronunciation. More natural pacing. Better emotional range.
My Italian digital instructor was explaining blood pressure management to Tamil-speaking patients with more clarity than I ever could.
Why This Actually Makes Sense
Here's what I learned after hours of debugging mumbled audio and truncated sentences:
Structure Over Sound
Multilingual AI models learn the structure of language, not just the sounds. They understand how syllables flow, where pauses belong, and how emphasis conveys meaning.
Native Speakers Have Habits
We rush. We swallow syllables. We assume the listener will fill in gaps. AI doesn't assume — it articulates.
Clarity Trumps Origin
The best voice isn't about origin — it's about clarity. A well-trained model speaking Tamil can outperform a native speaker who hasn't been trained in narration.
The Humbling Bit
I spent 29 years as a doctor, 12 of them as a GP in the NHS. I've explained complex medical conditions to thousands of patients. I thought I was a good communicator.
But when I heard Carmelo (the Italian) explain how the heart pumps blood in Tamil, I had to admit: he was better than me.
Not because he understood Tamil culture better. Not because he had some special connection to the language. But because the AI had been optimised for one thing: clarity.
And clarity, it turns out, trumps authenticity when you're trying to educate patients.
What This Means for Healthcare
We're entering an era where:
- A GP in London can create Tamil patient education that reaches clinics in Chennai
- A cardiologist in Mumbai can explain heart conditions in 10 languages they don't speak
- Health literacy is no longer locked behind language barriers
The AI doesn't replace clinical expertise. I still write the content. I still ensure medical accuracy. I still understand what my patients need.
But the AI amplifies my reach in ways I couldn't have imagined.
The Technical Bits That Made Me Laugh
After multiple failed attempts, here's what finally worked:
- Model: ElevenLabs Multilingual v2 (not the faster, cheaper Flash model)
- Key fix: Writing Tamil numbers as words, not compounds (ஐம்பது ஆயிரம், not ஐம்பதாயிரம்)
- Surprise learning: The AI was mumbling because I had written the text wrong, not because it couldn't speak Tamil
The AI was ready. I was the bottleneck.
My Takeaway
Pride has no place in learning.
If an Italian AI voice can teach my mother tongue better than me, I have two choices:
- Get offended
- Get curious
I chose curious. And now I have production-quality patient education that will help thousands of Tamil speakers understand their health conditions — in their own language, with crystal-clear audio.
My grandmother would be proud. Even if she'd laugh at me for needing a computer to speak Tamil properly.
For Fellow Clinicians and Healthcare Developers
If you're building patient education and hitting language barriers, here's what I learned:
- Don't assume native = better. Test everything.
- Use Multilingual v2 over Flash for quality (ElevenLabs)
- Write numbers as words in Indian languages
- Add periods between list items for natural pacing
- Listen to the full audio before publishing — truncation happens
The future of patient education is multilingual, AI-powered, and surprisingly humbling.
All hail the machines. 🙏
Topics: