Scaled Cognition Launches Voice TTS Model for Customer Service
Scaled Cognition has launched Voice, a text-to-speech model for customer service that it says reads account numbers, addresses and codes more accurately at lower cost, with API and self-hosted options.
Voice supports more than 25 languages, voice cloning and integration with existing AI agents, contact-center platforms and older telephone systems. Scaled Cognition said the product is intended to close a gap between speech that sounds natural and AI that reliably communicates details customers need to complete 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,” Akshay Kannan, vice president of product management at Scaled Cognition, said in emailed comments. Those difficulties are more pronounced in languages other than English, Kannan said.
The hosted API costs $9 per million characters, while self-hosting costs $1 per million characters, excluding the infrastructure needed to operate it. Kannan said one graphics processing unit can handle about 32 simultaneous conversations. A redundant setup using two GPUs on standard cloud hardware costs roughly $6,000 a month and supports about 500 million characters, he said.
Including infrastructure, Kannan estimated the cost at about $13 per million characters at that 500 million-character volume and less than $7 at 1.5 billion characters. He compared that with $35 to $100 per million characters charged by other hosted voices, which he declined to name. At low usage, self-hosting can cost more than the hosted API, the company said.
Scaled Cognition claims 70 milliseconds of speech-generation latency, measured from receiving text to starting speech rather than the total time an AI system takes to understand a caller and formulate an answer. Kannan said typical response time stays around 80 milliseconds even with 16 live conversations sharing a single GPU slice. He said distance is the biggest factor customers control and that running Voice in the same cloud region as the rest of a call stack adds only a millisecond or two.
Voice can interpret meaning before generating audio, the company said. Its Smart Pronunciation feature determines whether abbreviations refer to titles, street names or other information and how numbers should be spoken; for example, it reads “Pay 1/2 by 1/2” as “pay half by January second.” “Voice doesn’t just read text left to right,” Kannan said. “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 is Scaled Cognition’s model for customer-service conversations and actions in enterprise systems. When paired, APT supplies context that can help Voice resolve ambiguities beyond a single sentence, the company said. Voice does not require APT or another large language model, Kannan said; it streams audio through a standard API and is tuned for phone audio, allowing modern AI agents, contact-center platforms and legacy interactive voice response systems to use it. That compatibility could let enterprises update IVR systems without replacing underlying applications.
Scaled Cognition is offering a blind listening test comparing Voice with other models across price ranges. Kannan said early feedback favored Voice, but the supplied materials did not provide sample sizes or detailed results. The company has raised $100 million in Series A funding led by Khosla Ventures LLC and is positioning speech as another component of its enterprise AI platform.