Introducing Alto: the voice model driving a 5+ point lift in resolution

Last month at Decagon Dialogues, we introduced Voice 3, the first duplex voice agent handling customer calls in production. It combines an architecture that lets the agent listen, talk, and work at the same time with Chord, our speech model trained for customer conversations.

Today, we’re taking our voice agent further with Alto, a new model from Decagon Labs that powers the conversational layer of the duplex architecture. Alto determines when to speak and what to say. It knows when to answer a question, wait for confirmation, or keep the conversation moving while the agent takes action.

With Alto, our voice agent responds in about 700 milliseconds on average without relying on the generic filler statements that other systems use. On live customer calls, Alto has increased resolution rates by 5+ percentage points, with no other significant changes to the agent’s logic. This builds on the gains from Chord, which independently lifted resolution by up to 7 percentage points.

Hear it for yourself. Try our voice agent on the kinds of calls our customers handle every day.

"Customers have been trained for years to speak in short fragments at these IVRs just to get through to a human. The problem is, an LLM can’t do its job without real dialogue and context. But with Decagon the voice sounds like it’s actually listening, and it keeps the conversation moving instead of going quiet while it works. People start talking to it naturally without even thinking about it. That’s what Decagon delivers." - Christian Niedworok, Lead of Digital Service Communication at Deutsche Telekom

‍

The promise and tradeoff of duplex

Most voice agents process each turn in sequence through a cascade pipeline: transcribe, reason, then speak. A duplex agent splits that work across two models running in parallel: a fast conversational model that listens and speaks, and a more powerful reasoning model that handles complex decisions and actions behind the conversation.

Duplex brings interacting with voice agents much closer to talking with a person. Responses come faster. The agent keeps the conversation going while it works, instead of making callers wait in silence. And because part of the agent is dedicated to managing the conversation, it can be tuned to sound natural rather than like a system reading out results.

But getting a duplex agent to work well on real customer calls is hard. People interrupt, change their minds, and expect the right answer every time. For a conversation to feel natural, the agent has to be quick to respond, but smaller, faster models tend to make more mistakes.

That creates a decision on every turn: should the conversational model answer immediately or hand off to the slower reasoning model? Waiting can disrupt the flow, but answering too soon risks saying something that isn’t true, like telling a caller their reservation has been changed before the change is confirmed. The challenge is keeping the conversation natural without compromising reliability.

Alto makes conversations natural and reliable

Running our voice agent in production showed us where that tradeoff matters most, and we built Alto to address it. Its core skill is judgment: on every turn, it decides whether it can answer right away or needs to wait.

Alto continuously tracks what the caller has said and what the rest of the agent is working on. It uses that context to decide when to speak, give an update, or change course. If the agent needs a few seconds to look up information or make a tool call, Alto keeps the conversation moving, the way a person would on the phone. It grounds its responses in a shared record of the call, waiting for confirmation before telling a caller an action is complete.

Most voice agents fill the wait with, “Got it, one moment.” Alto puts that time to use. While a lookup runs, it can ask for the next piece of information it needs or check in with the caller. If the result comes back just as the caller asks something new, Alto answers the question and shares the result in the same reply, just as a human agent would.

How we trained Alto

Instructions alone can only take a model so far. To make Alto both natural and reliable, we trained it specifically for customer conversations.

We started with a small open-weight model to keep responses fast, then fine-tuned it on simulated calls informed by what we’d learned in production. Next came reinforcement learning. At each point in a conversation, the model generated several possible responses. We scored each one first on whether it correctly chose to answer or wait, then on whether its wording followed the business’s policies and reflected confirmed information.

The hardest part was finding examples that would help the model improve, because it already handled most situations correctly. We focused on cases where the right decision was less obvious and varied the policies and tools across simulated businesses. That helped Alto learn judgment it could apply across different situations, rather than memorize one company’s script.

Before any caller heard it, Alto was tested across more than 90 million simulated calls. It then ran in shadow mode alongside live production traffic, allowing us to evaluate its responses without having it speak to callers.

Specialized models for customer conversations

Alto is the latest model from Decagon Labs, and it follows the same principle as Chord: a model trained for one specific job will outperform a general-purpose model in customer conversations. Chord shapes how our agents sound; Alto shapes how they hold a conversation.

See how our agent handles your conversations here, or set up some time with us to learn more.