> ## Documentation Index > Fetch the complete documentation index at: https://docs.stenox.app/llms.txt > Use this file to discover all available pages before exploring further. # Meeting models and processing > Configure shared speech, local speaker separation, and independent insight providers. ## Three jobs, separate choices | Job | Current behavior | | - | - | | Speech to text | Uses the active dictation provider and model, including profile selection. | | Speaker separation | Local Nemotron where needed, or timed speaker labels from supported cloud speech adapters. | | Insights, refined notes, Ask | On-device Qwen 3.5 4B, or a selected subscription integration. | Quick setup's speech menu changes the shared speech model; there is no separate meeting-only speech override. A ready local speech runtime is reused. While a meeting holds it, competing jobs and model changes are guarded. Pause retains it; ending releases the meeting reservation without unloading dictation's selected speech model. ## Local meeting models **On-device** insights use **Qwen 3.5 4B** (`qwen3_5_4b_4bit`, catalog estimate \~3.1 GB). This is independent of the model selected for dictation Polish and uses a separate inference instance. For local speech and providers without speaker labels, Stenox loads **Nemotron-3-Diarization-8bit** for meeting-local speaker separation. It is not a dictation cleanup model and does not appear as an interchangeable speech model. Its state is scoped to one meeting and is released when that meeting finishes. Initial preparation may download these files. No fixed RAM minimum or processing speed is guaranteed. Local speech, diarization, and insights together need more resources than speech alone. Turning automatic insights off reduces the work while retaining transcription and personal notes. ## Cloud speech behavior Deepgram, AssemblyAI, and Gemini use timed meeting transcription with speaker requests. Other adapters use their normal file transcription path, with local speaker handling where needed. * **Deepgram** follows the selected Nova model and language. Streaming-only Flux uses Nova 3 for meeting audio windows, as it does for file transcription. Requests opt out of model improvement. * **AssemblyAI** uses Universal-3.5 Pro with its Universal-2 fallback and selected region/language. Stenox requests deletion of completed jobs. * **Gemini Transcribe** uses Gemini 3.5 Transcribe with Interactions storage disabled. These request settings do not promise zero provider retention. Cloud speech sends audio to that provider. Local windows are approximately ten seconds; cloud windows are approximately thirty seconds, with overlap. Transcription is not instantaneous and speaker identity may be uncertain across requests. ## Subscription insights In **Settings → Notetaker**, choose **Codex**, **Claude Code**, or **Google Antigravity**. Use **Sign in** if needed, then choose an available model and reasoning level. Models come from the connected runtime; Stenox does not promise a fixed list or silently replace a missing saved model. Codex and Claude Code use bundled runtimes; they do not require a separate Node, Bun, or CLI installation. Existing eligible coding-runtime sign-ins can be reused. These integrations require subscription authentication, not an API key. Provider limits and usage apply. Google Antigravity requires its official `agy` CLI installed separately. This adapter accepts 1.2 releases starting at 1.2.12. Its UI identifies the active CLI login source, not a verified email address. Selecting another insights provider or turning insights off does not sign you out of the external provider. Stenox does not expose a separate Sign out or Disconnect control for these subscription accounts. The session's **Insights model** menu also lets you change the provider, model, reasoning choice, and **Automatic insights** setting during capture or while editing saved history. Switching cancels pending insight work and rejects stale results while capture continues. Existing insights and your edits remain. Returning to On-device affects future work; it cannot recall earlier transfers. ## Context and failures Your selected provider handles Google-linked and manual sessions. With a cloud provider selected, the inputs below leave your Mac; **On-device** processes them locally. | Operation | Newly assembled input | | - | - | | Automatic reviews and Generate insights | Session title, bounded changed/recent conversation, speaker IDs and names, current outcomes, elapsed duration, new contribution IDs, included reference text, bounded selected-event context, and a Personal notes excerpt. | | Refined notes | Session title, whether to suggest a title, accepted visible outcomes with owner names, dates and status, included reference text, and a Personal notes excerpt. It does not independently add the transcript or full event snapshot. | | Ask Stenox | Your question, session title, bounded selected conversation with speaker labels, included references, and up to three prior question/answer pairs. It does not independently add Personal notes or the full event snapshot. | Personal notes guide focus in insights and refinement. Long notes use an excerpt based on a 6,000-byte UTF-8 limit plus an omission marker; the full saved note is unchanged. Notes do not establish facts, commitments, or owners. Editing or removing them invalidates stale results and refreshes the relevant provider context. Titles and saved content can contain information derived from [Google Calendar](/meetings/calendar). Codex and Claude Code can continue provider context across Insights, refinement, and Ask for the same session and selection; that context can contain Personal notes from earlier turns. Provider caching is not guaranteed. Antigravity sends the selected context afresh each time. A request's listed inputs do not imply that a continued provider session has forgotten previous inputs. Provider account terms, usage, retention, and training settings still apply. Selecting an integration does not establish a zero-retention or no-training guarantee. Invalid output, account errors, timeouts, or usage limits preserve recording and saved text. They may pause automatic reviews until an explicit retry. **Generate insights** remains an explicit action when automatic insights are off. There is no silent switch to another provider or paid API account. An audio capture or processing overrun can end a session; completed text is preserved. Check the status, export what is available, and resolve the input or resource problem before starting again. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.