Tool-Call Filler Audio

When your agent calls a real external tool — checking a calendar, sending an email or WhatsApp message, hitting a custom API — that tool takes real time to respond. Without anything to fill the gap, the caller hears silence for that entire stretch. Tool-Call Filler Audio closes that gap: if a tool call runs past a delay you configure, the agent speaks a short phrase ("One moment please...") in the caller's own language while it waits.

What triggers it

Filler only fires for tools that do genuine network work:

• HTTP API tools
• Integrations (Google Calendar, Gmail, WhatsApp, and other connectors)
• MCP tools

It never fires for the Calculator, End Call, or Transfer Call tools, or for the agent's own internal decision-making (like emotion detection) — those either run instantly or are control-flow, not something a caller would perceive as a delay.

Configuring it

Tool-Call Filler is a per-node setting, next to the node's Tools field:

1. Open the node where the tool is attached (Start Call or an Agent Node).
2. Turn on Speak Filler During Slow Tool Calls.
3. Set the Filler Delay — how long a tool call must be running before the filler speaks (100ms-3000ms, default 300ms). Keep this low enough that a genuinely slow tool call is caught, but high enough that a fast tool never triggers it.
4. Add a row under Filler Phrases for each language your agent speaks, with a comma-separated list of phrases for that language (e.g. One moment please..., Let me check that for you...). One phrase is picked at random each time, never repeating the same one twice in a row.

If the caller's active language doesn't have a phrase row configured, Talkr falls back to the workflow's primary language. If neither has phrases, no filler is spoken — silence is safer than a mismatched language.

How it behaves

• If the tool call finishes before the delay elapses, nothing is spoken — the filler only exists for genuinely slow calls.
• If a single agent turn triggers more than one slow tool call at once (a batched function-call turn), only one filler is spoken for that batch, not one per tool.
• Filler speech follows the node's own Allow Interruption setting, exactly like any other agent speech — it's not specially protected from being interrupted, and it's not specially uninterruptible either.
• Filler speech isn't added to the conversation's LLM context — it's a UX cue for the caller, not something the model needs to reason about.

Not available on realtime agents

Realtime (speech-to-speech) providers — OpenAI Realtime, Gemini Live, Grok Realtime, Ultravox — generate audio directly from the model and don't have a separate text-to-speech step. Tool-Call Filler hooks into that TTS step, so it has no effect on realtime agents.