A voice stack for any Agent, QuixiVoice runs on Metal first
QuixiVoice uses Rust + Burn to string VAD, streaming ASR, Gemma, TTS, and an MCP server into one voice stack. Currently it only runs Metal kernels; CUDA, CDNA, and XPU are on the porting roadmap.
QuixiVoice has announced a voice stack for any Agent. Eric Hartford breaks it down clearly in a tweet: a fast voice server plus an MCP server, built underneath with Rust and Burn.
For ordinary users, getting a voice Agent running usually means stringing together four things: VAD determines whether someone is speaking, ASR converts speech into text, the language model understands and generates a reply, and TTS reads the reply out loud. QuixiVoice's choices are Silero VAD v5, NVIDIA Nemotron 3.5 ASR Streaming 0.6B, Google Gemma 4 E2B IT (QAT), and Ming 3.0 TTS 0.5B. The first two handle listening, Gemma handles thinking, and Ming handles speaking. Reference voices come from LAION's 6k-diverse-reference-voices dataset, aiming to cover more diverse voices.
The MCP server is what sets this apart from an ordinary voice demo. Agents can access voice capabilities through MCP, requiring less glue code. The difficulties with voice Agents often lie in pipeline assembly and tool integration; individual models are only part of the equation. Building the server and MCP together makes this engineering-wise closer to a foundation that can be used directly.
Demo video attached to the tweet:
It currently runs on Metal kernels, i.e., the Apple GPU path. Eric Hartford says he will port the kernels to CUDA, CDNA, and XPU. Based on this information, QuixiVoice is currently more of a runnable starting point; cross-hardware support is still on the way. The tweet also tags @dnhkng and @ChristophSchuh6. What is worth watching next is porting progress and actual latency data.
Tweet link: https://x.com/QuixiAI/status/2122167653456388222
发布时间: 2026-11-16 11:25