Scaled Cognition launches cost-effective voice model for enterprise customer service

Scaled Cognition Inc., which serves enterprise customer service operations, unveiled Voice today, a text-to-speech model engineered to pronounce account numbers, addresses and similar details with greater precision than alternatives while charging significantly less. The offering ships as an API or as a self-hosted option. It covers more than 25 languages, includes voice cloning functionality and integrates with existing AI agents, contact center platforms and legacy phone systems. Voice cloning generates a synthetic version of a person's voice.

The product addresses a specific market need: balancing natural-sounding output with the reliability required when customers need to confirm transactions. "Modern generative TTS models can sound fantastic, but they often stumble over the details that matter most in customer service, like reading back an alphanumeric confirmation code, a dollar amount or a street address," said Akshay Kannan, vice president of product management at Scaled Cognition. The challenge intensifies for languages beyond English, according to Kannan. Scaled Cognition contends that Voice delivers both naturalness and accuracy without the premium pricing of competitors.

Pricing and infrastructure economics

Voice carries a price tag of $9 per million characters via the hosted API or $1 per million characters for self-hosted deployment, though the latter excludes infrastructure expenses. Kannan broke down the cost structure for self-hosting. A single graphics processing unit manages roughly 32 concurrent conversations. A redundant configuration with two GPUs on standard cloud infrastructure runs approximately $6,000 monthly and processes about 500 million characters, he noted.

When factoring in infrastructure, Kannan projected costs of roughly $13 per million characters at 500 million characters monthly and below $7 at 1.5 billion. He contrasted these figures with $35 to $100 per million characters from other hosted voice services, which he did not identify by name. "Self-hosting gets economical faster than people expect," he said.

Deployment flexibility

At minimal usage levels, self-hosting expenses can surpass the hosted API rate. "Most of the enterprises we work with already run their own infrastructure, so Voice just drops in," Kannan noted. "For teams that don't want to manage it, we host it for them."

The company also highlights 70 milliseconds of speech-generation latency—the interval between receiving text and beginning audio playback, distinct from the total time an AI system requires to comprehend a caller and generate a response. "Even with 16 live conversations sharing a single GPU slice, typical response time stays around 80 milliseconds," Kannan stated. "The biggest factor customers control is distance." Placing Voice in the same cloud region as the rest of a call infrastructure introduces only one or two additional milliseconds. "That's a big reason we let enterprises self-host it right next to their agents and telephony," Kannan said.

Smart Pronunciation and context awareness

A key differentiator of Voice is its capacity to extract meaning before generating speech, the company stated. The Smart Pronunciation feature determines whether abbreviations denote titles, street designations or other categories and how numerals should be articulated. For instance, it renders "Pay 1/2 by 1/2" as "pay half by January second."

"Voice doesn't just read text left to right," Kannan explained. "Before it says a word, Smart Pronunciation works out what the sentence actually means, powered by the same customer-experience-tuned intelligence behind our APT model." APT represents Scaled Cognition's model for customer-service conversations and operations within enterprise systems. When combined, APT supplies conversational context that assists Voice in resolving ambiguities beyond individual sentences.

Voice operates independently of APT or any other large language model. "It streams audio through a standard API and is tuned for phone audio, so anything that can send text and play back speech can use it, from modern AI agents and contact center platforms to legacy [interactive voice response systems]," Kannan said. This compatibility could enable enterprises to modernize interactive voice response systems without overhauling the underlying infrastructure.

Competitive positioning

Scaled Cognition is conducting a blind listening test comparing Voice against competing models across various price tiers. Kannan reported that preliminary reactions favored Voice, though the materials provided did not include sample sizes or comprehensive findings.

The company, which secured $100 million in Series A funding led by Khosla Ventures LLC, is presenting voice capabilities as an additional layer of its enterprise AI platform.