Speaker Layout

The Speaker Layout node frames a whole recording by speaker in one step. Nothing is cut, so the video is as long as the recording. With one camera, each speaker is framed in turn. With several cameras of one conversation, each speaker is shown on their own camera.

Inside, the node runs the same steps that you can wire by hand: Transcribe with speaker labels and word timings, then Speaker View over the whole recording. Speaker View puts a box on each speaker on screen and switches to whoever is talking. Each step runs as its own job, so a long recording, a cancel or a failure behaves as it does in a hand-built workflow.

Speaker Layout renders with Speaker View, which runs on Nodaro Cloud. On a self-hosted install, connect the install to Nodaro Cloud, under Integrations › nodaro.ai or with an API key. Without a connection, the node is not offered and a run is refused before anything starts.

When to use it

• Make a vertical, speaker-framed version of a whole two-person video podcast.
• Make a multicam interview in which each speaker is on their own camera, without cutting anything.
• Make a square or 4:5 version of a panel recording, with every speaker in a box.

To cut the recording too, for example its silences and filler words, use Tighten Episode with a framing. For short clips, use Clip Pack.

Quick start

Inputs

These are refused before anything is charged:

• A video that is already narrow. With one recording, a video that is as narrow as the chosen shape, or narrower, has no extra width to follow a speaker across. Choose a narrower shape, or use a wider video.
• A recording with no picture. There is nothing to frame.
• A recording that cannot be read.

With more than one recording wired into Recordings, the first one is Camera 1. Drag the recordings in the settings panel to reorder them.

Outputs

Settings

The crossfade switch is always greyed, with the reason "Nothing is cut out of the recording, so there is nothing to crossfade." Pan and Zoom are greyed under Side by side, Stacked and Grid, which keep each speaker's place fixed. A greyed switch that a node or an API request sets anyway is drawn as a cut.

Each speaker on screen gets an automatic box. A voice that is heard only briefly, under 10 seconds and under 5% of the talk, gets no box of its own. A stray label in the transcript then never takes a share of the frame.

Following faces

Framing source sets where the speakers' boxes come from:

• Automatic (follow faces). Follows faces while Speaker Frames is available to you, and uses the even split otherwise, with no refusal and nothing extra charged. The estimate follows the same rule.
• Even split. Each speaker on screen gets an even share of the frame.
• Follow faces. Speaker Frames tracks each face, and Speaker View follows it. Speaker Frames is charged per started block of 1,000 sampled frames. Faces are not matched to speakers yet: a camera that shows one person is followed, and a camera shared by several speakers keeps the even split.

Where Speaker Frames is not available to you, the settings panel offers no framing source, and Automatic (follow faces) is the even split. A node set to Follow faces then shows a note, and its run is refused with 403 node_not_available before anything is charged.

Several cameras

Wire each camera's recording into Recordings. When you have one, wire the mixer or recorder track into Master audio. The run then lines the recordings up with Audio Sync, transcribes the sound, and works out which camera each speaker is on with Speaker Map. With two or more cameras, Camera Switch puts each stretch on whoever is talking. Speaker View then frames each camera's own shot.

• Recordings. Up to six recordings of one conversation, in order: Camera 1, Camera 2 and so on. At most five of them can be cameras. The sixth place is for the sound: a Master audio or one sound-only recording.
• Master audio. The video's sound, and the clock that the recordings are lined up to. Without a Master audio, the node picks one sound for the whole video, so the sound never jumps between camera microphones.
• Wide camera. The camera that shows everyone. A speaker with no camera of their own is shown on it. There is none by default.
• Recordings already in sync. Turn it on when every recording already shares one clock. The run then skips the Audio Sync step.
• Who is on which camera. By default, Speaker Map decides it from the cameras' own microphones. Set a speaker to a camera, or to the wide shot, to override it. Speaker 1 is the first voice heard, Speaker 2 the next, and so on.

The run's result names who was put on which camera, with the notes of the line-up and of the map. Where Speaker Map is not available to you, a run in the editor or in an app reads only the first recording, and the settings panel says so.

Review before render

With Review before render on, a run stops at a preview. Speaker View renders it at its lower preview rate. Watch the preview, then click Render final on the node. It renders the video at full quality, with the settings the preview was made with.

• There is nothing to edit. The layout has no cut to change, so the node has no review editor.
• Render final renders only the render. Transcribe, Speaker Frames and, with several cameras, Audio Sync, Speaker Map and Camera Switch are not run or charged again. There is no fee.
• Someone has to review. A workflow or app run with nobody to review it cannot stop at the preview, and fails with preview_review_required. The MCP tools speaker_layout and continue_use_case are reviewed by the AI assistant that calls them.
• Not once per list item. A node with Review before render on that is fed by a List is refused before any of its runs starts, with nothing charged. With it off, a List runs the node once per item, each as its own run with its own fee.

If a run fails

The steps that finished stay finished. Their results, such as the transcript, the line-up and the camera map, are kept, and they are not charged again. Press Continue on the node: only the step that failed and the steps after it run again. A failed run's fee is refunded, and Continue holds it again, so a run is charged its fee once.

Credits

Speaker Layout charges each step at its own rate, the same as the node it stands for, plus a small per-run fee. The node shows the full estimate, fee included, before you run.

• The per-run fee is charged once per run. A Render final after a review is not charged the fee again.
• A List runs each item as its own run, with its own fee.
• You are charged for the framed video's own minutes, not for the estimate.

The estimate before you run

The estimate prices Speaker View at the full length of the recording, which is the length of the video it makes. Because the layout is not known before the run, the estimate uses Speaker View's higher rate. A recording with at most two speakers on screen and no zoom is charged the lower rate. A recording with no known length is estimated at the 180-minute ceiling. With several cameras, Speaker View is priced at the length of the sound: the Master audio, or else the longest recording.

Tips

• Keep Automatic framing. Automatic (follow faces) follows faces where it can, and costs nothing extra where it cannot.
• Review long recordings. Turn on Review before render and check the preview before you pay for the full-quality render.
• Reuse the edit. Wire the Edit (EDL) output into a second Speaker View to render the same framing in another shape.
• Cut and frame. To remove silences and filler words as well, use Tighten Episode with a framing.

Troubleshooting

From the API

POST /v1/speaker-layout starts a run and answers 202 with jobId and executionId. The body takes mediaUrls, a list with one video recording or up to six recordings of one conversation, and optionally masterAudioUrl, cameras, targetAspect, layout, switchType, emphasisStyle, framingSource and reviewBeforeRender. Poll the job with GET /v1/jobs/:id/status.

• cameras takes wide (a recording number, or null), synced (true when the recordings already share a clock) and speakerCameras, which maps a speaker number to a recording number. 0 means the wide shot.
• framingSource is "auto", "static" or "tracking". Left out, it is "auto". emphasisStyle is "none", or "scale", "border" and "dim" joined with "+".
• The job's output_data carries the framed video (videoUrl), its length in seconds (durationSec) when the run knows it, the edit (json), the transcript and quality (proxy for a preview, final otherwise). A run on several recordings adds cameras, the report of who is on which camera. A failed run names the failed step in failedStep.
• An Idempotency-Key header makes a retry of the same request answer with the same jobId and "deduped": true, and start nothing.
• POST /v1/speaker-layout/estimate takes the same body and answers { data: { totalCredits, stepsCredits, feeCredits, durationSec } }, without starting anything.
• POST /v1/speaker-layout/continue with { jobId, mode } continues a run. "mode": "from-failed" runs what a failed run did not complete. "mode": "render-final" renders a review run's final. It takes no editedEdl: one sent is refused with 400 use_case_no_plan.

The SDK (client.edit.speakerLayout, client.edit.estimateSpeakerLayout, client.edit.continueUseCase), the CLI (nodaro edit speaker-layout, with --estimate for a quote, and nodaro edit continue-use-case) and the MCP tools speaker_layout, estimate_use_case and continue_use_case call the same routes. See Run a single node, Editing in the SDK and the MCP tools.