Use an AI assistant in an NDI® control room with Dual AI Studio | Apex

In an NDI® control room, Dual AI Studio provides a second technical and presentation perspective using images and transcribed speech. Start by checking inputs, describing your programme and running a manual analysis before enabling an automatic schedule. The team retains the decision over which advice is used on air.

Choose the role and prepare inputs

Tech focuses on technical feedback; Perf on presentation; DUAL uses both channels. They share model configuration with separate instructions and outputs. This does not mean two different models analyse every video frame. This workflow covers Dual AI Studio 1.9, the version Arnaud Frey reports testing on macOS Tahoe and Golden Gate.

1. In INPUTS / OUTPUTS (ENTRÉES / SORTIES in the French screenshots), choose the camera or NDI video source and check the preview.
2. Select the microphone or NDI audio source separately. Video and audio can come from two different sources: speak and check the meter.
3. In Intelligence & context, describe the programme, presenter’s role and useful feedback. Adapt the News, Gaming or Interview examples.
4. Enable VISION to include the image in the analysis workflow. Check transcription mode too; on-device processing requires the relevant macOS settings, including Dictation.
5. Choose the mode and click MANUAL ANALYSIS (ANALYSE PONCTUELLE), or LAUNCH DUAL AI (LANCER DUAL AI) in DUAL. Review the result before enabling voice outputs.

Configure Ollama or Gemini

In API SETTINGS (API RÉGLAGES), choose the provider. For Ollama, start the server and install the required model: selecting it in the menu does not install it. http://localhost:11434 refers to this Mac; another address sends requests to that other machine. Use Test connection, save, then run an analysis: a successful connection does not confirm that the model responds correctly.

For Gemini, enter your key, choose a model accessible to your account and use Save. Names illustrated in the manual are examples. Gemini uses a remote service; review the Gemini API terms applicable to your account, region and service before submitting sensitive content.

Schedule analysis and listen to voices

AUTO-COACHING (cadence) triggers periodic analysis. TECH and PERF sliders set images analysed per minute, not video frames per second. A busy channel can skip a trigger. Reduce frequency if responses accumulate or arrive too late to help.

In inputs and outputs, set TECH VOICE (VOIX TECH), PERF VOICE (VOIX PERF), then AUDIO OUT 1 (TECH) and AUDIO OUT 2 (PERF). Begin with headphones or a separate output so voices do not return to the microphone. If integrating them into studio monitoring, provide a simple mute and priority for human communications; these outputs alone are not an IFB intercom or mix-minus.

Distinguish windows, NDI text, audio and programme

TECH, PERF and PRES floating windows display text on the Mac. Opening them does not automatically overlay them on video. In NDI, enable TECH, PERF and PRES text outputs separately from TECH/PERF audio outputs. PGM forwards the active video and audio input: check “Dual AI Studio - PGM” on the receiver.

To change a locked input, stop the activities using it: recording, streaming and PGM. Check each destination you use rather than inferring its behaviour from the preview alone.

Common mistake: treating Ollama as an entirely local session

Check the whole chain: Ollama address, effective transcription mode, additional services and logs. Transcription may use remote processing if local processing is unavailable. Google Live is separate; its Dev/DEBUG controls are not guaranteed in a distributed release. When present and enabled, check its effect on image transmission independently of VISION. The Ollama FAQ helps distinguish its local and cloud uses.

Logs may contain context, transcripts and responses. Clearing the console display does not necessarily delete the files. Review exports before sharing. To select a configuration, begin with a non-confidential rehearsal, compare useful and incorrect advice and measure its delay in your actual workflow.

NDI® is a registered trademark of Vizrt NDI AB.