Straight answers.

Transcription and speaker separation do, completely — they use on-device sherpa-onnx models with no network. The only step that can reach out is the AI assistant, and only to the provider you choose. Point it at a local Ollama or LM Studio server and even that stays on your machine.

It reliably separates you from everyone else — a direct voice and audio from a loudspeaker are acoustically far apart. Telling two remote people apart through one shared loudspeaker is much harder, and the app is honest about that rather than inventing confident-looking labels. If the remote side is one person, it’s a non-issue; if it’s several, you can rename and merge speakers, and tune the match threshold.

Overlapping speech gets a single label. Splitting overlap needs a separate audio channel per speaker, which a single microphone doesn’t provide — so it’s out of scope by design, not a bug.

Claude (Anthropic, the default), OpenAI, Google Gemini, a local Ollama or LM Studio server, or any other OpenAI-compatible server — llama.cpp and vLLM locally, or OpenRouter, Groq, DeepSeek and others hosted. You choose in Settings and can switch anytime.

Only for the hosted providers. For a local server there’s no key at all — the app just asks the server which models it has loaded. The rest of the app (capture, transcript, speaker separation) needs no key and no account whatsoever.

It writes the summary, decisions, and action items — and hangs supporting detail under the bullets you type. But it never touches the words you wrote yourself: your bullets appear in the document exactly as typed. That separation is structural, not a prompt the model might ignore.

The default speech models are English. sherpa-onnx offers multilingual streaming models, and the model paths are configurable in Settings if you want to swap them in.

Not today — Acousma ships as a Windows 10/11 (x64) desktop app. The engine is written to be portable, so other platforms are possible down the line, but there’s no build for them yet.

The app itself is free. If you use a hosted AI provider you pay that provider directly for the assistant calls, under your own key. Run a local model and there’s no per-call cost at all.