Scaled Cognition targets customer service with low-cost voice model

Scaled Cognition is taking a shot at a very specific pain point in customer service: voice software that sounds polished but mangles the details people actually need. Its new text-to-speech model, Voice, is meant to read account numbers, addresses, confirmation codes and similar information more reliably than current systems, while also keeping the bill down.

The product is available either through an API or as a self-hosted deployment. It supports more than 25 languages, includes voice cloning, and is designed to plug into existing AI agents, contact center platforms and older telephone systems. That matters because the company is not asking enterprises to rip out what they already use. It wants to slide in beside it.

Pricing is the main pitch. Scaled Cognition says Voice costs $9 per million characters through the hosted API, or $1 per million characters if customers self-host it, not counting infrastructure. Akshay Kannan, the company’s vice president of product management, said a single GPU can handle about 32 simultaneous conversations, and a redundant two-GPU setup on standard cloud hardware runs about $6,000 a month while covering roughly 500 million characters.

On that math, Kannan said the effective cost comes out to about $13 per million characters at that 500 million-character level and below $7 at 1.5 billion. He also put competing hosted voices at between $35 and $100 per million characters, though he didn’t name them. At low usage, though, self-hosting can cost more than Scaled Cognition’s hosted API, so the cheaper option depends on volume.

The company is also leaning hard on pronunciation intelligence. Its Smart Pronunciation feature tries to figure out whether abbreviations are titles, street names or something else before speaking, and the company says that lets Voice turn “Pay 1/2 by 1/2” into “pay half by January second.” Voice can work on its own, but it also pairs with Scaled Cognition’s APT model, which provides customer-service context when the two are used together. The company says early blind listening tests favored Voice, though it didn’t share the sample size or the results in detail.