Decagon Voice 3 ships with Chord on the call path
Decagon Voice 3 launched Oct 1 with Chord and a duplex call stack. Unite.AI and Business Wire agree on the product list; benchmarks stay company-reported.
- Unite.AI's October 1 feature and a Business Wire-syndicated Dialogues release both date the launch to October 1 and name Voice 3, Chord, Personal Agent Gateway, Agent Modules, and Duet Apprentice.
- Both describe a duplex architecture: a low-latency conversational layer beside a heavier reasoning and tool-calling layer.
- Unite.AI alone walks Chord's training claims and the company-reported resolution-rate lifts; treat those lifts as Decagon's numbers, not a third-party audit.
in this block
Decagon Voice 3 landed on October 1, 2026 at Decagon Dialogues with Chord, the company's first in-house speech model, and two independent write-ups plus the company's own Dialogues framing agree on the product shape even when their evaluation numbers refuse to collapse into one score.
What actually happened
Decagon Voice 3 is the customer-call agent. Chord is the voice it speaks with. Unite.AI says Product Manager Quique Lores and Senior Product Marketing Manager Ariana Xiang authored the Voice 3 announcement, while co-founders Jesse Zhang and Ashwin Sreenivas framed the wider Dialogues slate. The Business Wire write-up, carried on regional financial wires, repeats the four-product list and the duplex pitch without becoming a second lab notebook. That is enough for a confirmed product launch. It is not enough to treat every benchmark chart as gospel.
Chord's design pitch is specific. Unite.AI says the model was post-trained for customer-experience conversations rather than audiobook narration, that it slows for confirmation codes and phone numbers, and that it supports more than 70 languages with mid-sentence switching. The Business Wire piece also highlights pace changes for confirmation codes and the duplex idea that conversation and task execution run in parallel. Shared claims stay shared. The tokenizer-free diffusion detail and Modal serving note stay with Unite.AI's deeper technical summary of Samuel Zhang's companion post.
Deutsche Telekom and Chime quotes appear in Unite.AI's recounting of the announcement. Christian Niedworok's line about years of IVR training customers to speak in fragments, and Janelle Sallenave's line about brand customization with cross-channel memory, are company-provided customer quotes. They are not independent lab results.
Evaluation numbers that stay attributed
Unite.AI reports Decagon's customer switches in telecom, financial services, and travel: resolution-rate gains of 6.1, 7.1, and 2.6 points when only the voice changed, plus barge-rate drops of 14.5, 6.1, and 5.3 points. A Zeniteq summary opened in research echoed those company-reported lifts and warned they lack independent verification. This article keeps them as Decagon-reported. It does not promote them to peer-reviewed fact.
Blind-test language is messy across Decagon's own surfaces. Unite.AI's Voice 3 piece summarizes one announcement line as about 90% of users unable to tell, while the technical companion path describes a test where 45.7% of listeners thought the real human was the AI. Those are not the same sentence. memcool does not average 90 and 45.7. Preference-test figures such as 36.2% overall across 185 listeners also stay inside Unite.AI's account of Decagon's materials.
Training claims without the brochure voice
Unite.AI's technical path says Chord uses a verbatim transcription process that keeps pauses and filler words, plus a recording method that mixes script content with freer delivery. It also says Decagon post-trained a tokenizer-free diffusion model because tokenizing audio throws away detail. Those are company technical claims relayed by Unite.AI. The Business Wire release does not re-litigate the diffusion stack. If you are a speech engineer, that gap is your cue to read Samuel Zhang's companion post directly rather than trust a secondary paraphrase.
Licensed data and consented voice talent, never customer-owned call audio, is another Unite.AI-relayed Decagon line. That matters for enterprises that cannot feed production calls into a vendor's training loop. It is not a third-party compliance certificate.
Where it sits in the agent stack
Personal Agent Gateway is the sibling product for when a customer's own agent shows up. Agent Modules stretch journeys past support. Duet Apprentice learns from internal resources and escalations. Readers comparing other October agent control planes can open the OpenClaw Enterprise note. Readers watching decision-model spam week can open the Clef decision models note. Neither is a phone agent with Chord.
Primary sources: Unite.AI's Voice 3 feature and the Business Wire Dialogues release as carried on myMotherLode's wire page.
What to do as a reader
If you evaluate Decagon Voice 3, separate the shipped architecture claims from the company-run scoreboard. Ask for the duplex demo and the language switch under interruption. Do not buy a "90% human" slogan without reading the companion test design. Nothing here is a vendor recommendation. Not financial advice. DYOR, ser.
Not financial advice. DYOR, ser.
