QuietVoice: voice control for AI agent swarms

QuietVoice is an asynchronous voice control plane for headless, distributed AI coding agents. You manage a swarm of remote agents by voice from your phone: the agent speaks its question to you in a cloned voice as a Telegram voice note, and you answer out loud — QuietVoice hands it back to the agent not as a raw transcript but as structured intent + emotion. The system is two planes joined by one contract. A tiny always-on control plane (CGO-free Go, no GPU) exposes two MCP tools — say and listen_voice — so any agent (Claude Code, Codex, ...) can speak and listen. A swappable inference plane, "inferenced", keeps the models hot and is byte-identical from a Mac to a consumer GPU to an AMD MI300X; only the engine binaries are rebuilt, the Go layer is the same across CUDA and ROCm. Accuracy comes from Gemma. Speech models transcribe the words in parallel, then Gemma reconciles them against the raw audio AND against the agent's own question — so it doesn't just hear the words, it knows what you are answering. A dictation box cannot do that. Self-hosted on AMD, this is our best-use-of-Gemma case. Meaningful AMD use is measured, not guessed: replication only time-slices one GPU (~2x ceiling anywhere), while the datacenter card unlocks what replication structurally cannot — continuous batching, 11.28x on MI300X. Same stack, swap binaries: portability measured at 5.3x from an M1 to the datacenter. The full listen/speak plane runs ROCm-native on AMD. Open source (MIT). Runs on ROCm.