AI Voices Just Got 4x Cheaper to Run in the Cloud

Text-to-speech (software that reads words aloud in a human-like voice) usually runs in the cloud, on rented servers billed by the second. The problem is that even when nothing is happening, those servers quietly cost money — an idle voice app can hold onto gigabytes of memory doing nothing. Two researchers, Pakorn Nathong and Kunat Pipatanakul, set out to fix that waste rather than chase faster or smoother audio.

Their fix uses two tricks. First, each spoken request gets a fixed, small slice of processing power instead of grabbing everything at once, so requests stop fighting each other when traffic spikes. Second, the system releases its memory after quiet periods but keeps the core program warm and ready — like a restaurant that sends servers home during slow hours but leaves the kitchen fired up. The moment a new order arrives, it snaps back to work.

The numbers are striking. Using a small open-source voice model called Kokoro-82M, their setup produced 2.71 seconds of audio for every second of computer time, versus 0.89 seconds with standard tools. Cost per hour of audio dropped from $0.0631 to $0.0153 — about 4.1 times cheaper. Idle memory use fell from 8.7 GB to 1.33 GB, and the system started speaking in 2.2 seconds instead of 7.7 seconds after a cold start. Notably, all of this ran on ordinary CPUs, no pricey graphics chips required.

Why should you care? Cheap voice generation means more products can afford natural-sounding speech: audiobooks, screen readers for blind users, language-learning apps, automated phone menus and voice assistants. When a cost drops by 75%, features that were too expensive to offer become routine. The honest caveat: this is a six-page technical report, not a shipping product. The gains were measured under specific conditions, and real savings depend on your traffic patterns and how your cloud provider bills. It also doesn't make voices sound better — just cheaper.