One GPU, one job
Even when hosted, a single GPU serves one narration at a time. A 5-minute script takes about 6–8 minutes to generate.
The pipeline needs an NVIDIA GPU with ~4 GB of free memory for every second it is generating. Free hosting tiers don’t offer GPUs, and a GPU that’s always on costs money every hour, even when nobody is using it.
| Option | Cost | GPU | Result |
|---|---|---|---|
| Free web hosts (Render, Fly, Railway…) | Free | ❌ | Runs out of memory loading PyTorch; CPU-only even if it fit |
| Hugging Face Spaces (free) | Free | ❌ CPU only | Minutes per chunk; long scripts impractical |
| Oracle Cloud Always Free | Free | ❌ ARM CPU | Same problem: far too slow |
| My laptop + a free tunnel | Free | ✅ RTX 3050 | Works great, only while the laptop is on |
| Rented cloud GPU, always on | ~$0.3–1 per hour | ✅ | Hundreds of dollars a month for a portfolio demo |
| Serverless GPU (pay per use) | Per second | ✅ | Viable for a product; cold starts of ~40 s to load the model |
One GPU, one job
Even when hosted, a single GPU serves one narration at a time. A 5-minute script takes about 6–8 minutes to generate.
Voice cloning needs guardrails
A public cloning service invites misuse. Consent checks, rate limits and abuse handling become a product of their own.
Instead of a demo that would be offline most of the time, this site shows real outputs, real measurements and the reasoning behind them, and it’s always online because it’s just static files.