Tavus Sparrow-2 turn-taking vs writing pauses in a Sume clip

Tavus exposes turn-taking as settings on a live agent: turn_detection_model (sparrow-1 or sparrow-2), turn_taking_patience and pal_interruptibility, each low, medium or high. A Sume avatar clip has no turns to manage, so pacing is authored: a pause is a video_inputs scene with voice.type: "silence" and a duration you choose.

If you are comparing the two for a sales or support flow, the question is whether the viewer will talk back. If they will, you need turn-taking. If they will not, you need good pacing.

The Tavus settings

Tavus lists sparrow-1 and sparrow-2 as the turn detection models. The page gives sparrow-1 as the default and says the default changes to sparrow-2 on September 8, 2026; it calls Sparrow-2 the frontier model and lets you opt in by setting turn_detection_model. turn_taking_patience runs from low (quick replies that may cut in on natural pauses) to high (waits for a clear end of turn), with medium as the default. pal_interruptibility sets how readily the agent stops when the user speaks (Tavus docs: Conversational Flow). Because the page ties the default to a date, check it for the model your account gets today.

Pacing in a rendered clip

In a Sume multi-scene avatar video each entry in video_inputs has a voice. A spoken scene uses type: "text" with a script; a non-speaking beat uses type: "silence" with a required duration and no script (Generate avatar video). The total planned duration must still land inside the 4-60 second window.

Write the pause where a live agent would wait for an answer: after a question to the viewer, before a price, after a surprising number. A two second silence scene reads as the agent giving the viewer time to think.

• Keep pauses short. In a 30 second clip, one or two beats of 1-3 seconds is usually enough.
• Silence scenes count toward the planned duration, so a script near the 60 second limit has less room for beats.
• Captions are built from the spoken text, so a silence scene adds a gap rather than words.

Which approach fits

Pick live turn-taking when users reply. Pick authored pacing when you control the whole message. Griffin, which Tavus describes as a full-duplex model that continuously decides when to respond, is offered only as a restricted Griffin-Lite research preview for select testers, and Tavus says it is not available to customers at this time, so today's live option is the Sparrow-based pipeline.

A worked pacing example

Take a 30 second clip that asks a question. Write a spoken scene of about 12 seconds that sets up the problem, a 2 second silence scene, a spoken scene of about 10 seconds with the answer, a 1 second silence, and a closing line of about 5 seconds. The total stays under the 60 second ceiling with room to spare.

Preview the result before a long run. Because beats are authored, you can change a duration and render again rather than retuning a model setting, and the change is exactly what you asked for.