Enterprise customer service vendor Scaled Cognition Inc. is introducing a text-to-speech model today that it says provides more accurate pronunciation of account numbers, addresses and other details at a lower cost than competing services.

The product, called Voice, is available through an application programming interface or for self-hosted deployment. It supports more than 25 languages, voice cloning and can work with existing artificial intelligence agents, contact center platforms and older telephone systems. Voice cloning creates a digital replica of a given person’s voice.

The launch targets a gap between natural-sounding speech and AI that reliably communicates information 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 e-mailed comments.

Those difficulties become more pronounced in languages other than English, Kannan said. Scaled Cognition said Voice enables enterprises to get natural-sounding speech and accurate readings without paying premium prices.

Voice costs $9 per million characters through the hosted API or $1 per million characters for self-hosting. The latter figure excludes the infrastructure needed to operate it.

Kannan explained how the math works out to lower costs than competing providers. One graphics processing unit can handle about 32 simultaneous conversations using Voice. 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 costs at about $13 per million characters at that 500 million-character rate and less than $7 at 1.5 billion. He compared that to between $35 and $100 per million characters charged by other hosted voices, which he declined to name.

“Self-hosting gets economical faster than people expect,” he said.

Cloud option

At low usage, the self-hosting cost can exceed Scaled Cognition’s hosted API price. “Most of the enterprises we work with already run their own infrastructure, so Voice just drops in,” Kannan said. “For teams that don’t want to manage it, we host it for them.”

The company also claims 70 milliseconds of speech-generation latency, the time from receiving text to starting speech, rather than the total time an AI system takes to understand a caller and formulate an answer.

“Even with 16 live conversations sharing a single GPU slice, typical response time stays around 80 milliseconds,” he said. “The biggest factor customers control is distance.” Running Voice in the same cloud region as the rest of a call stack adds only a millisecond or two of latency. “That’s a big reason we let enterprises self-host it right next to their agents and telephony,” Kannan said.

One of the distinctive features of Voice is its ability to interpret meaning before generating audio, the company said. For example, its Smart Pronunciation capability 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 the company’s model for customer-service conversations and actions in enterprise systems. When the products are paired, APT supplies conversation context that can help Voice resolve ambiguities beyond an individual sentence.

Voice doesn’t require APT or another large language model to operate. “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.

That compatibility could give enterprises a way to update interactive voice response systems without replacing the underlying applications.

Scaled Cognition is also offering a blind listening test that compares Voice with other models across price ranges. Kannan said early feedback favored Voice, although the supplied materials did not provide sample sizes or detailed results.

The company, which raised $100 million in Series A funding led by Khosla Ventures LLC, is positioning speech as another component of its enterprise AI platform.

Image: SiliconANGLE/Microsoft Designer