Decagon launches Voice 3 with its Chord speech model
Decagon pairs Voice 3 with Chord and a duplex architecture for customer calls. Here is what changed and what its early results establish.
Decagon announced Voice 3 on October 1, 2026, pairing its customer service voice agent with Chord, the first speech model from Decagon Labs. The Voice 3 launch focuses on keeping conversation flowing while an agent completes a task.
The release was part of Decagon Dialogues, alongside Personal Agent Gateway, Agent Modules and Duet Apprentice. Together, the announcements extend the company’s customer service platform into more channels and business workflows.
A voice agent that listens while it works
Decagon describes a duplex architecture with two layers operating in parallel. A fast conversational model handles listening and speaking. A separate, more capable model handles reasoning, tool calls and guardrails. The architecture announcement says this lets the agent answer questions and provide progress updates during longer tasks.
The agent also processes incoming audio while speaking. According to the launch overview, it can continue through a brief acknowledgment such as “mhm” while yielding to a genuine interruption. That distinction matters when a caller wants to correct information without restarting the interaction.
Chord brings a dedicated model to customer calls
In its technical explanation of Chord, Decagon says it further trained a diffusion model that represents speech without audio tokenization. Its data pipeline preserves pauses, changing emphasis and conversational hesitations. The goal is speech that adapts to the moment rather than reading every sentence with the same delivery.
The Voice 3 announcement says Chord uses licensed data and voice talent who gave consent, excluding customer owned data. Delivery can slow for details such as confirmation codes.
Decagon’s voice product page lists support for more than 70 languages, automatic language switching, customizable voices and transfers to human agents with a summary. These are advertised platform capabilities. The announcement does not provide a language by language evaluation of Chord.
The evidence and access limits
Decagon reports that three customer programs saw resolution rates rise by 2.6 to 7.1 percentage points after switching voices to Chord. Its results report describes comparisons before and after the change, with only the voice altered. These are vendor reported customer outcomes, and the post does not give sample sizes for those comparisons.
The company also reports a 36.2 percent overall preference share in a comparison involving Chord and three other speech models, based on 185 listeners. That listening test measures stated preference and does not establish accuracy or reliability on live calls.
The public access route is a Decagon demo request. The launch posts do not publish standalone Chord API access, downloadable weights or pricing. Prospective customers still need to confirm deployment terms and costs with Decagon.
For teams evaluating Voice 3, the useful test is a full service call. Measure task completion, interruptions, response delays and transfers to people under realistic conditions before treating a natural voice as proof of reliable service.
