Voice agents learned to sound human. Now they are learning when to shut up

The judgment layer

Decagon has launched Alto, a new model from Decagon Labs that powers the conversational layer of its duplex architecture. Alto determines when to speak and what to say during customer conversations, and the Decagon voice agent responds in about 700 milliseconds on average with Alto. Alto has increased resolution rates by 5+ percentage points on live customer calls, and the Chord model independently lifted resolution rates by up to 7 percentage points. m:…5444

The architecture is the insight. Duplex agents split work across a fast conversational model and a powerful reasoning model, so the agent can keep talking naturally while a stronger model does the thinking. Decagon introduced Voice 3 last month at Decagon Dialogues, and Voice 3 combines an architecture that allows simultaneous listening, talking, and working with the Chord speech model. Alto was trained using reinforcement learning on simulated calls, and was tested across more than 90 million simulated calls before production. m:…5444

The metric that matters here is resolution, not realism. A voice agent that sounds human but cannot finish the call is a demo. A model that knows when to speak, when to wait for reasoning, and when to hand off is a product. The 5+ point lift is the company's claim about its own model on live calls, so treat it as a vendor number, but the direction is the story: the competition moved from voice quality to conversational judgment. m:…5444

The interaction layer

ElevenLabs is building the same layer from the platform side. ElevenLabs introduced new models, agents, and deployment options at its NYC Summit, including the Interaction Engine, which manages conversational rhythm and turn-taking, and selects between cheaper and frontier models based on task complexity. The Memory feature allows ElevenAgents to recall customer details across interactions, and Routines enable multi-step automation before, during, and after conversations. m:…7474

ElevenLabs is also pushing voice into the enterprise workflow. ElevenAI is a new app for enterprise teams to generate notes and action items, powered by the Scribe v3 model. ElevenLabs is partnering with McKinsey and QuantumBlack, and is developing a domain-optimized model for the financial services industry. ElevenAgents is used by Yahoo, Westgate, and KPN. Mati Staniszewski predicts conversational agents will become critical enterprise infrastructure, which is a prediction, but the partner list is a fact. m:…7474

The plumbing underneath all of this is getting quieter and faster. Quail Voice Focus 2.3 is a primary-speaker isolation model that sits in front of STT or S2S systems, and keeps the most dominant voice of those it hears, even when that speaker is quiet, distant or in a reverberant room. Version 2.3 reduces deletions compared to version 2.2, accepting more insertions as a trade-off. The model uses 16-bit weights, which halves the file sizes compared to previous versions, and new SIMD kernels process each audio frame 27-38% faster on x86-64 CPUs. The models are language-agnostic and do not require separate settings per language. m:…4651

Translation is the hard case

Real-time translation is where the judgment layer meets its hardest test. LiveKit Agents 1.8.11 introduced beta support for OpenAI's gpt-realtime-translate and Google's gemini-3.5-live-translate-preview. Interpreter models are optimized for fluency and low latency rather than verbatim accuracy, which can lead to errors in critical domains like healthcare, and they do not accept system prompts, tools, or mid-stream steering. m:…5524

The latency stacks up. Bridged calls using interpreter models can accumulate 300ms to 800ms of latency per exchange, and chained translation pipelines can add 2 to 4 seconds of latency per direction. Interpreter models lack in-line speaker attribution in synthesized audio, and text-level evaluation is insufficient for testing interpreter models because failure modes are audio-shaped. Roark provides simulation and scoring across 45 languages and accents, and interpreter models update behind the API, necessitating production call replay to detect drift. m:…5524

Put together, the week in voice is about the same question from several directions: when does the agent speak, what does it remember, whose voice does it keep, and how do you test it when the failure is in the audio rather than the text. The companies answering those questions are the ones building the interaction layer, and the interaction layer is where the voice agent business will be won. m:…5444 m:…7474 m:…4651 m:…5524