Voice AI's full-duplex era: xAI goes sub-second at $0.08 a minute

The product: sub-second voice at a metered price

xAI announced Grok Voice 3.0, a speech-to-speech model with "class-leading latency at under one second," available in its Voice Agent API at $0.08 per minute. The pricing is the tell: voice AI is now sold like telephony, by the minute, and the competition is on latency and benchmark scores. m:…5590

xAI's claimed scorecard: #1 on the Artificial Analysis Speech-to-Speech Index at 86.9, the top accuracy on ServiceNow's EVA-Bench for enterprise voice agents at 74%, 97% on Full Duplex Bench (pauses, turn-taking, interruptions), and 71.6% on τ-Voice — "comparable with GPT-Live-1 (Astra)". m:…5590

The deployment claim is the most striking: xAI says Starlink's call centers resolve roughly 70% of customer calls with no human in the loop using Grok Voice. That's a support-cost story and a revenue story in the same sentence — which is exactly why voice agents are getting funded. m:…5590

The research: turn-taking without the tax

Full-duplex conversation lives or dies on turn-taking: knowing when the caller is done without awkward pauses or constant interruption. JAL-Turn, in a preprint, proposes a lightweight, speech-only framework that fuses acoustic and linguistic cues through cross-attention. The engineering insight: share a frozen ASR encoder so turn-taking prediction runs in parallel with speech recognition, adding almost no end-to-end latency. It is designed to run as a small parallel task rather than a job for the big model. m:…2795

This is the unglamorous half of the voice race. The model can be brilliant, but if turn-taking costs a second of compute, the conversation feels broken. JAL-Turn's approach — treat it as a small parallel task, not a job for the big model — is the kind of optimization that makes sub-second voice economically viable. m:…2795 m:…5590

JAL-Turn also ships what practitioners need most: a scalable pipeline for automatically deriving reliable turn-taking labels from large-scale real-world dialogue corpora. The bottleneck for turn-taking models has been labeled data, not architecture — automating the labels is what makes the lightweight approach deployable. m:…2795

The architecture detail that matters: JAL-Turn uses a cross-attention module to adaptively fuse pre-trained acoustic and linguistic representations for turn-taking prediction, and sharing a frozen ASR encoder lets it run in parallel with speech recognition with almost no additional end-to-end latency. m:…2795

The open alternative

Fish Audio released Drama 3 as open source: a 35B-parameter mixture-of-experts backbone that activates only 3B parameters per token, paired with a hybrid linear-attention backbone and a new audio decoder. It decodes a single stream at 271 frames per second on one H200. m:…1460

The feature list reads like a dubbing studio's wishlist: multi-turn revision, scene dubbing, editing by infilling, and streaming text input, all in one model. Open weights won't match xAI's end-to-end latency story, but for media production — where the job is directed performance, not real-time conversation — an open 35B MoE with infilling changes who can build a voice studio. m:…1460

Drama 3's training recipe is doing a lot of work: three-stage pre-training, instruction fine-tuning, then large-batch reinforcement learning "whose only reward is a single learned model of human preference". Its 271 frames per second on one H200 compares against 177 for its predecessor S2 — a generational jump in throughput for open speech models. m:…1460