Scaled Cognition targets customer service with low-cost voice model

Enterprise customer service vendor Scaled Cognition Inc. has unveiled a new text-to-speech model named Voice, promising accurate pronunciation of critical details like account numbers and addresses at a more affordable price than competitors. The Voice product is accessible via an API or can be self-hosted, supporting over 25 languages and offering voice cloning to create digital replicas of individual voices.

Scaled Cognition's Vice President of Product Management, Akshay Kannan, highlighted that while modern generative TTS models produce high-quality speech, they often struggle with accurately conveying essential information customers need to complete transactions. The Voice model excels in this regard, particularly for languages other than English, according to Kannan.

Enterprises can obtain natural-sounding speech and precise readings at a lower cost with Voice, priced at $9 per million characters for the hosted API or $1 per million characters for self-hosted deployment. The self-hosting option excludes the cost of necessary infrastructure, which Kannan estimates at about $13 per million characters when supporting 500 million characters, and less than $7 at 1.5 billion characters.

This pricing is notably lower than the $35 to $100 per million characters charged by other hosted voices. Voice boasts a speech-generation latency of 70 milliseconds, significantly faster than traditional AI systems. By running Voice in the same cloud region as an enterprise's existing call stack, latency further decreases to around 80 milliseconds, even when 16 conversations share a single GPU slice.

A unique feature of Voice is its Smart Pronunciation capability, which interprets meaning before generating audio, ensuring accurate pronunciation of abbreviations, numbers, and other critical details. This feature relies on Scaled Cognition's APT model, which optimizes customer-service conversations and actions within enterprise systems.

Kannan emphasized that Voice does not require APT or another large language model to function, making it compatible with various existing systems, including modern AI agents and legacy interactive voice response systems. The company is also conducting blind listening tests to compare Voice's performance against other models across different price ranges, with early feedback favoring Voice.

Scaled Cognition, having raised $100 million in Series A funding led by Khosla Ventures LLC, views speech as an integral part of its enterprise AI platform.