Decagon introduced Voice 3 and its Chord speech model for customer calls. The new design runs conversation alongside a separate model handling reasoning and tools, so a lookup can continue while the caller asks another question. Chord adjusts its delivery within a sentence. Decagon says its training uses licensed material and consenting voice talent, excluding customer-owned data.
Decagon, a developer of customer-service AI agents, introduced Voice 3 and its Chord speech model on 1 October. Voice 3 uses a duplex architecture, allowing listening and speaking to continue while the agent performs work in the background. The product is designed for customer calls, where a follow-up question can arrive before a lookup or other task has finished.[1], [2]
A conversational model handles listening, speech and progress updates. A separate, more powerful model manages reasoning, tool calls and the rules constraining those actions. Decagon describes a call in which the agent can handle a smaller request while a longer task remains active. Its announcement contrasts this with a sequence of speech recognition, language-model processing and speech synthesis, where coordinating turns becomes difficult as each stage adds delay.[1]
Chord, the first speech model developed by Decagon Labs for customer conversations, changes delivery phrase by phrase. A confirmation code or phone number can be read more slowly before the voice returns to its conversational pace. Teams retain controls over the voice and the pronunciation of business terms. Decagon says training uses licensed data and consenting voice talent, with customer-owned data excluded.[1]
