Backchannel

Let the agent naturally acknowledge the caller while they speak.

Introduction

Backchanneling (“filler words”) makes the agent sound natural and attentive. When enabled, the bot occasionally interjects short phrases like “mhm”, “I see”, “right” while the caller speaks - exactly like a human listening on the phone.

How it works

The agent follows both sides of the conversation and interjects the way a person does: at a natural break, not on a timer.

1. When. A nod lands on a micro-pause (the caller went quiet for a moment but has not finished) or at the end of a sentence or clause in the live transcript. Only when neither a pause nor punctuation shows up for a while does the agent fall back to a timed nod.
2. Whether at all. The agent stays silent when the caller:
   • has spoken for less than the minimum speech duration or said only a few words (a short answer, not a story),
   • asks a question or checks the line (“hello?”, “are you there?”),
   • is answering a question the agent just asked - until the answer turns into a longer explanation,
   • has just interrupted the agent (a few seconds of silence, so the “mhm” is not heard as the bot trying to keep the floor),
   • and always while the agent itself is speaking.
3. What. The phrase follows what the caller is saying right now:
   • dictating a number, e-mail, address or spelling a name → a quiet “mhm”, “uh huh”,
   • describing a problem or a complaint (“doesn’t work”, “never arrived”) → “I see”,
   • enumerating (“and then”, “apart from that”) → “mhm”, “uh huh”,
   • otherwise mostly a short “mhm”, longer words less often. The same phrase is never used twice in a row. The agent only ever uses the phrases from the settings list - if, say, “I see” is not on it, it picks from what it has.
4. The audio is injected into the stream without interrupting the caller and without involving the LLM.
5. A minimum gap is kept between consecutive interjections.

Short breathing pauses do not reset the caller’s turn: speaking time and word count run from the caller’s first word until the agent replies.

Configuration

Go to Agent → Settings tab:

Setting	Description	Default
Enabled	Turns backchanneling on/off	Off
Phrases	Comma-separated list of phrases	Language defaults
Probability	Chance (0–1) of an interjection at each opportunity	0.6
Min. speech duration	Seconds of caller speech before the first interjection	3.0
Gap	Minimum gap between interjections (s)	5.0

Content recognition covers Polish, English, German, Spanish, French and Italian. In other languages the agent still nods at breaks, but without matching the phrase to the content.

Default phrases per language

Language	Phrases
Polish	mhm, rozumiem, aha, ok, tak, jasne
English	mhm, I see, uh huh, okay, right, got it
German	mhm, verstehe, aha, okay, ja, genau
Spanish	mhm, entiendo, ajá, vale, sí, claro
French	mhm, je vois, ah oui, d’accord, oui, bien sûr
Italian	mhm, capisco, ah sì, ok, sì, certo

Best practices

• Short phrases: 1–2 syllables work best (“mhm”, “uh huh”, “ok”).
• Don’t overdo it: high probability + short gaps sound artificial.
• Brand consistency: match the phrases to the agent’s personality.
• Language culture: use acknowledgments natural for the region.

Technical details

• Sounds are generated by TTS ahead of time and cached.
• Backchannel uses the same voice as the agent.
• It doesn’t interfere with STT transcription or the LLM.
• It works alongside noise reduction and VAD settings.

Limitations

• Interjections don’t appear in the transcript (they’re injected past it).
• Too-frequent short phrases can sound repetitive.
• Requires correctly working VAD. Micro-pauses are detected best with Smart Turn enabled; without it the agent relies on punctuation from the transcript.
• Matching the phrase to the content is simple dictionary rules, not a language model - on purpose, since a nod has to land within a fraction of a second.