Scaled Cognition Introduces Advanced Text-to-Speech Model for Enterprises
Enterprise customer service provider Scaled Cognition Inc. has launched a text-to-speech (TTS) model named Voice, which promises to deliver enhanced accuracy in pronouncing essential details such as account numbers and addresses, all while providing a more cost-effective solution compared to competing services.
Available through an application programming interface (API) or for self-hosted deployment, Voice supports more than 25 languages, features voice cloning, and integrates seamlessly with existing AI agents, contact center platforms, and legacy telephone systems. Voice cloning technology allows for the creation of a digital replica of a person’s voice.
The product aims to bridge the gap between natural-sounding speech and the reliable communication of critical transaction information that customers need.
“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, in an email statement.
Kannan noted that these challenges are especially pronounced in languages other than English. He asserted that Voice equips enterprises with an opportunity to achieve both natural-sounding speech and precise readings without incurring premium costs.
The pricing structure for Voice is $9 per million characters for the hosted API option and $1 per million characters for self-hosting, excluding the necessary operating infrastructure. Kannan elaborated on the cost efficiency, explaining that a single graphics processing unit (GPU) can accommodate approximately 32 simultaneous conversations using Voice. A redundant configuration utilizing two GPUs on standard cloud hardware translates to a monthly cost of around $6,000, supporting nearly 500 million characters, he noted.
Cost-Effective Solutions for Businesses
Incorporating infrastructure, Kannan estimated the costs could reach about $13 per million characters at the 500 million-character rate and drop to less than $7 at 1.5 billion characters, in stark contrast to the fees ranging from $35 to $100 per million characters charged by other hosted TTS providers, which he did not specifically identify.
“Self-hosting becomes economically viable faster than many anticipate,” he stated.
For organizations with lower usage, self-hosting costs may surpass Scaled Cognition’s hosted API pricing. “Most of our enterprise partners already maintain their own infrastructure, making Voice a simple addition,” Kannan explained. “For those who prefer not to manage this, we offer a hosting option.”
Scaled Cognition claims a speech-generation latency of 70 milliseconds—the time taken from receiving text to initiating speech—noting that the total time for an AI system to process a caller’s input and generate a response is longer. “Even with 16 live conversations sharing one GPU slice, average response times remain around 80 milliseconds,” said Kannan. He emphasized that distance is a critical factor; hosting Voice in the same cloud region as the call infrastructure adds negligible latency, hence the company’s decision to allow enterprises the option to self-host alongside their telephony systems.
A unique aspect of Voice is its ability to interpret context before producing spoken audio. Its Smart Pronunciation capability discerns the intended meaning of abbreviations and determines the appropriate pronunciation of numbers. For instance, it interprets “Pay 1/2 by 1/2” as “pay half by January second.” Kannan highlighted this capability, stating, “Voice doesn’t simply read text left to right. Before verbalizing, Smart Pronunciation deduces the true meaning of a sentence, leveraging the same customer experience-focused intelligence that powers our APT model.”
APT represents Scaled Cognition’s model designed to enhance customer-service interactions and actions within enterprise frameworks. When used in conjunction with Voice, APT provides conversational context, aiding in resolving ambiguities that extend beyond singular sentences.
Voice can operate independently from APT or other large language models. “It streams audio through a standard API and is specifically optimized for telephonic audio, thus compatible with any system capable of sending text and playing back speech, ranging from modern AI agents and contact center solutions to older interactive voice response systems,” Kannan clarified.
This compatibility offers enterprises an opportunity to update their interactive voice response systems without the need for complete application replacements.
Additionally, Scaled Cognition is conducting a blind listening test to compare Voice against various models across different price ranges. While Kannan mentioned that early feedback has been favorable for Voice, he did not disclose specific sample sizes or detailed results.
Scaled Cognition, which recently secured $100 million in Series A funding led by Khosla Ventures LLC, is setting forth to enhance its enterprise AI platform with this new speech technology.