Decagon launches Voice 3 and Chord, its own speech model for customer calls

Decagon/Decagon agents

Voice 3 runs two models side by side: a fast one that listens and talks, and a stronger one that reasons, calls tools and enforces rules. Chord, the first model from Decagon Labs, is a speech model post-trained on customer-service calls that slows down for things like confirmation codes.

• In a blind test, listeners picked the human voice correctly only about 10% of the time.
• 70+ languages, switching mid-sentence when the caller does.
• Chord trained on licensed data and consented voice actors, not customer data.
• Keeps talking during lookups and gives progress updates on slow tasks.