Scaled Cognition, a lab that builds AI models for customer-service agents, released a text-to-speech model called Voice on 20 October 2026. It is on sale from day one, either through a hosted API or as software a company runs itself, and it is meant to slot into existing AI agents, contact-centre platforms and older phone menu systems.
“Self-hosting gets economical faster than people expect”
Akshay Kannan, vice president of product management, Scaled Cognition, to SiliconANGLE · siliconangle.com ↗
The headline is the price: $9 per million characters hosted, or $1 per million self-hosted, not counting the hardware. The model covers more than 25 languages, can clone voices, and starts speaking about 70 milliseconds after it receives text. That figure leaves out the time an agent spends deciding what to say. The company puts one GPU at roughly 32 simultaneous conversations. By its sums, a redundant two-GPU setup at about $6,000 a month handling 500 million characters works out near $13 per million all-in, and under $7 at 1.5 billion. It says rival hosted voices charge $35 to $100 per million, without naming them.
A telephone handset beside a studio microphone, illustrating the speech launch
TokenPost.ai · Image: TokenPost
Scaled Cognition argues that other speech models sound good but trip over the details a support call turns on: confirmation codes, amounts and addresses, and more so outside English. Voice includes a text-normalisation layer that decides how to read abbreviations, dates and account numbers, so that "St." can come out as "Saint" or "Street" depending on context. Brand names and specialist terms can be tuned. Its evidence for the quality claim is a blind listening test of five pairs of customer-service clips against other providers, with no sample sizes or detailed results released.
About Scaled Cognition
Scaled Cognition is an AI lab that builds models for enterprise customer-service agents, aiming for answers verified against a company's own policies. Its flagship model before Voice was APT, short for Agentic Pretrained Transformer.
The company announced a $100 million Series A, led by Khosla Ventures with Genesys joining, on 25 June 2026; the Wall Street Journal put its valuation at about $750 million. Co-founder and CTO Dan Klein is a UC Berkeley AI professor who earlier co-founded an agentic AI company that was sold to Microsoft.
Voice does not depend on the company's APT conversation model, or on any language model. Used together, APT supplies the context to resolve ambiguity and Voice reads out the verified details. The company raised a $100 million Series A led by Khosla Ventures, and CMSWire names founders Dan Klein and Dan Roth. SiliconANGLE points out that self-hosting can cost more than the hosted API at low volumes, and more languages are planned. The company describes Voice as the first speech model built specifically for enterprise customer experience, a claim it makes for itself rather than one the coverage tested.
Sources · 3