Scaled Cognition, an AI lab building models for enterprise customer service agents, introduced Voice, a text-to-speech model, SiliconANGLE's Paul Gillin reported on October 20. The company says Voice reads account numbers, addresses and similar details more accurately, and at lower cost, than competing services.
What the model does
Voice supports more than 25 languages and voice cloning and works with existing AI agents, contact center platforms and older telephone systems. A capability called Smart Pronunciation interprets meaning before generating audio, such as deciding whether an abbreviation is a title or a street name. It does not require the company's customer-service model, APT, or another large language model, though pairing the two lets APT supply conversation context. Akshay Kannan, vice president of product management, said modern generative TTS models can sound fantastic but often stumble over the details that matter most.
Pricing and performance claims
The hosted API costs $9 per million characters, and self-hosting costs $1 per million characters excluding infrastructure. Including infrastructure, the company estimates about $13 per million characters at 500 million characters and under $7 at 1.5 billion. It says competing hosted voices charge $35 to $100 per million characters, without naming them. One GPU can reportedly handle about 32 simultaneous conversations, with speech-generation latency of 70 milliseconds. A blind listening test showed early feedback favoring Voice, but no sample sizes or detailed results were provided. Kannan said: “Self-hosting gets economical faster than people expect,” and: “The biggest factor customers control is distance.”
Availability and funding
Voice is available through an API or for self-hosted deployment. The company raised $100 million in Series A funding led by Khosla Ventures.
Source: SiliconANGLE. This is an AI Business Watch summary of the original report.