On 12 November 2026 Anthropic rolled out Claude Meetings in beta. Claude can now join a Google Meet call as a visible participant, follow the conversation, answer questions out loud or in chat (replies stay in writing by default), build docs and whiteboards that teammates edit live, and drop notes plus action items afterward. Support is Google Meet only for now; other platforms are promised. It is limited to Team and Enterprise plans, stays out of calls with external guests unless an admin allows it, and leaves if one joins. It keeps a transcript and screen descriptions, not the raw audio.

Basic waveforms illustrating the raw signal before engine processing

That is useful for teams that already live in Meet and want an agent that starts work while people are still talking. It does not solve the problem of a developer or interviewer who needs a local, low-latency overlay that hears both sides of any call, turns the remote audio into text in another language, and optionally surfaces a short translated reply or suggestion without leaving the Windows desktop.

Why the cloud participant does not move the first millisecond

Claude Meetings runs on Anthropic’s infrastructure. The Meet client ships the media; Claude consumes the resulting transcript and context. The round-trip for a spoken question plus a written answer is whatever the network and model queue deliver. For many knowledge-work meetings that is fine. For a live language overlay that must stay under a few hundred milliseconds end-to-end so the listener can still follow the original speaker, the local path dominates:

  1. Endpoint and mode — USB headset, Bluetooth HFP/A2DP, built-in, or exclusive-mode WASAPI.
  2. Dual-track acquisition: one capture client for the local mic, one render-loopback client so the remote party’s audio stays distinct and timed.
  3. Windows Audio Engine (audiodg.exe + audioeng.dll) in shared mode. Mixing, communications enhancements and vendor APOs routinely add 10–40+ ms and become non-deterministic under load.
  4. Consumer thread: lock-free SPSC ring copy, format conversion, resampling, VAD.
  5. Streaming STT (local GGUF or remote) plus optional local model polish for translation or short replies.
  6. Local TTS writing into a render client if a spoken reply is wanted.
  7. Overlay paint on a dedicated thread that never blocks the audio callback.

Bluetooth HFP alone can inject another 50–150 ms of buffering that no remote agent recovers. Even if Claude finishes a note in under a second, the overlay still waits for the Windows path that delivered the samples.

A minimal shared-mode capture callback that stays under 1 ms and does no allocation looks like this:

// WASAPI shared-mode capture — callback must be allocation-free static HRESULT STDMETHODCALLTYPE CaptureCallback(IAudioCaptureClient *pCapture) { UINT32 packetLength = 0; BYTE *pData = NULL; DWORD flags = 0; HRESULT hr = pCapture->GetBuffer(&pData, &packetLength, &flags, NULL, NULL); if (FAILED(hr) || !packetLength) return hr; // Writer side of a pre-allocated power-of-two SPSC ring (float32 frames) ring_push_float32(capture_ring, (const float *)pData, packetLength); pCapture->ReleaseBuffer(packetLength); return S_OK; }

The consumer owns resampling (libsamplerate or a fixed polyphase), VAD, and the write to the model or WebSocket. Keep the audio thread pure; any lock or malloc here shows up as glitches or higher DPC latency.

Labeled sine wave showing amplitude and trough

Practical trade-offs that still matter

  • Cloud agent vs local stack: Claude Meetings is excellent when everyone is already in Meet and the goal is collaborative artifacts. A local dual-track pipeline is required when the call can be any app (Zoom, Teams, Discord, a raw SIP client) and the user needs the remote audio turned into text in a different language before the next sentence finishes.
  • Shared vs exclusive mode: Shared mode is convenient and survives device changes, but the engine and APOs add jitter. Exclusive mode gives tighter periods at the cost of exclusive device ownership and more brittle device switching.
  • Local vs remote STT: A 45 ms cloud final is excellent when the network is clean. Local models trade higher first-token latency and GPU/CPU load for privacy and offline use. Hybrid is common: local for the hot path, cloud for higher accuracy on finalized turns.
  • Overlay thread isolation: The paint must never contend with the audio callback. A dedicated UI thread reading from a second lock-free queue keeps glitches out of the speakers.

Claude Meetings is a strong product for the meetings it targets. It does not replace the need for a careful Windows capture and render path when the goal is a live language overlay that works on any call.

I'm building Talkaside, a Windows app for job interviews and calls in another language.
Drafted with AI assistance from this project's code and logs.