IVR Voice Configuration — Settings, Prompt Tricks, and Troubleshooting

This guide covers how to configure your IVR bot's voice, behavior, and speech settings. It also explains the difference between IVR-specific settings and the channel overrides, and includes tricks for improving how the bot greets callers and reads information aloud.

---

What Is the IVR Voice Configuration?

Each phone number connected to Velaro has a Voice Configuration that controls:

• Which voice the bot uses (ElevenLabs or Polly)
• How quickly it responds after the caller stops speaking
• When to transfer to a live agent
• What to say during idle moments or at the end of a call
• Pronunciation rules and background noise handling

These settings are configured per phone line in IVR Settings → Voice Configuration.

---

Key Settings Explained

Speech Detection (Deepgram)

Setting	What It Does	Default	Notes
End of Speech Timeout	Seconds of silence before the bot responds	3.0s	Increase for slow speakers; decrease for fast back-and-forth
Silence Duration	Milliseconds of silence before speech is considered complete	2500ms	Lower = faster response; higher = waits longer for pauses
Interrupt Sensitivity	How easily a caller can interrupt the bot mid-sentence	Low	Low = less accidental interruption; High = very responsive
Speech Model	Deepgram transcription model	nova-3-general	flux saves ~1.5s per turn using confidence-based detection (see below)
EoT Confidence Threshold	Deepgram confidence cutoff for end-of-turn (flux model only)	—	Set to 0.6 when using flux; ignored for nova-3-general
Min Confidence	Transcription confidence threshold below which input is ignored	0.50	Lower = processes more speech but may catch noise

Faster Responses with the Flux Speech Model

The default speech model (nova-3-general) uses silence to detect when a caller has finished speaking — it waits for 2–3 seconds of quiet before responding. This is reliable but adds latency to every turn.

The flux model uses Deepgram's confidence score instead: it responds as soon as it's confident the caller has finished their sentence, regardless of silence duration. This typically saves 1.5–1.8 seconds per turn on clear phone calls.

When to use flux:

• Clear phone environments (office lines, mobile calls indoors)
• Callers who speak at a normal pace
• Any line where response speed is a priority

When to stay on nova-3-general:

• High-background-noise environments (restaurants, outdoor, call centers)
• Callers who pause frequently mid-sentence (elderly callers, hesitant speakers)
• If callers report getting cut off

How to enable: Edit the voice config → Speech model dropdown → select "flux" → EoT Threshold: 0.6 (recommended). Save and test with a real call.

Filler Phrases (Thinking Out Loud)

When the flux model or any speech model is in use, the bot can play a short natural phrase while it's computing a reply — preventing awkward dead air. A phrase like "One moment." or "Let me check on that." plays after a brief delay (default 400ms). If the bot responds faster than the delay, no phrase is played.

Customizing filler phrases: In the voice config editor, you can enter your own list (one phrase per line). Leave it blank to use the system defaults. If your brand voice is formal, try phrases like "One moment please." or "Let me look that up." Avoid phrases like "Absolutely!" or "Certainly!" — they sound scripted.

Disabling fillers: Set Filler Delay to 0 to turn them off completely.

Default phrases (system uses these when none are configured): "One moment.", "Let me check on that.", "Sure, just a second.", "Got it, give me a moment.", "Let me pull that up.", and others — chosen randomly, never repeating the same phrase back to back.

Voice (ElevenLabs)

Setting	What It Does	Default
Voice ID	The ElevenLabs voice used for text-to-speech	Configured per line
Speed	Speaking pace — 1.0 = normal	1.0
Stability	How consistent the voice is — lower = more expressive variation	0.6
Similarity	How closely the voice matches the original — lower = more natural variation	0.85
Model	ElevenLabs TTS model	turbo_v2_5

Tip on Stability: If the bot sounds flat or monotone, lower the stability value (try 0.4–0.5). This allows the voice more natural variation. If it sounds inconsistent or "wobbly," raise it back toward 0.6–0.7.

Idle & Timeout

Setting	What It Does	Default
Idle Check-In (seconds)	How long the caller is silent before the bot checks in (10 to 120)	20s
Idle Check-In Message	What the bot says at the check-in. Leave blank for the standard "Are you still there?"	Standard message
Idle Hangup (seconds)	How long after the check-in, with no answer, before the bot says goodbye and ends the call (5 to 120)	20s
Max Call Duration	Hard cap on total call length	30 min

These are set in each phone line's voice configuration. If the caller starts speaking during the idle goodbye, or right after it, the hangup is cancelled and the bot carries on.

---

Escalation Options

You can configure which escalation paths are available when a caller asks to speak with a person:

• Bridge Call to Agent — Transfers the caller directly to a phone number. Requires Enable IVR Bridge Call to be on in your subscription and a Fallback Transfer Number set in IVR Settings for the line. Without that number the bot cannot transfer.
• Offer Callback — Bot takes the caller's phone number and arranges a return call.
• Take Message — Bot collects name, issue, and callback number; sends a notification to your team.

You can enable or disable each option individually in Voice Configuration → Escalation.

If the bot says "live agent transfer is not available" but you have a transfer number configured: Contact support — the Enable IVR Bridge Call subscription flag may need to be turned on for your account.

---

IVR vs. Email — These Are Separate Settings

A common source of confusion: the IVR channel has an option called "Suggest emailing documents to callers" in Channel Overrides. This teaches the voice bot to offer "Would you like me to email that to you?" during a phone call. It does not affect how the bot behaves in actual email conversations.

Email conversations (when someone contacts you through your email inbox) are configured separately under the Email channel in Channel Overrides.

These two settings are independent and do not sync — because a customer may want the voice bot to offer email delivery but keep email-thread behavior at default, or vice versa.

---

Improving Your Greeting — Pacing and Tone Tricks

The welcome greeting is the first thing a caller hears. Here are tricks to make it sound more natural:

Add Natural Pauses

Use ... (ellipsis) between phrases where you want a beat:

Hello, and thank you for calling Sullivan Tire and Auto Service... your place for tires and auto service. How can I help you today?

The ... causes a brief natural pause in the voice output. Use it for emphasis or between distinct thoughts.

Add Energy to a Section

Ending a phrase with ! signals the voice engine to add more energy to that sentence. This works in the Welcome Greeting because that text goes directly to the TTS engine.

Important: Do not put anything in parentheses in the Welcome Greeting — the voice engine reads parenthetical text aloud, so (with warmth) would be spoken as the words "with warmth."

Tone Hints in the Bot Prompt (Not the Greeting)

If you want to guide the bot's tone throughout a call — not just at the opening — add a stage-direction hint to the bot's AI prompt in AI Settings → Bot Configuration:

When greeting a caller, speak with warmth and genuine enthusiasm — vary your rhythm so it doesn't feel scripted.

These instructions are read by the AI and shape how it phrases responses. They are never spoken aloud.

Read Phone Numbers Clearly

The system automatically instructs the bot to read phone numbers with natural pauses between digit groups:

> 2 0 7... 7 7 4... 8 4 7 3

If you find a number in your KB data that's being read too fast, check that it's formatted with hyphens: 207-774-8473. Hyphens help the TTS engine place natural pauses.

Never Asks "Do You Want Me to Give You That Number?"

The bot is instructed to give information directly without asking permission first. If you're seeing this behavior, it may mean the bot's AI configuration prompt is overriding this behavior — check the bot's system prompt under AI Settings → Bot Configuration.

---

Voicemail & Email Handling on IVR

The IVR bot will never read an email address aloud over the phone — it says "the email we have on file" instead. This is intentional: callers can't easily write down email addresses during a phone call.

When the bot confirms contact information for a callback or follow-up, it will say:

> "I have your email address on file — is that still the best way to reach you?"

It will not say the actual address. This is by design.

URLs from your knowledge base are also never read aloud — the bot will offer to send the link via SMS or email instead.

---

How Calls End

The bot does not hang up just because a caller says "thanks" or "okay" after an answer. It asks "Before you go, is there anything else I can help you with?" and keeps the call open. The call ends only when the caller signals they are done (for example "no, that's all" or "goodbye"), when a workflow reaches its end, or when the idle hangup fires.

Grace window: After the bot's goodbye there is a short pause (about 2.5 seconds) before the line closes. If the caller starts talking during the goodbye or right after it, the hangup is cancelled and the bot continues the conversation. This also applies to the idle-timeout goodbye. A workflow that has reached its end still ends the call.

If a call stays open longer than expected: The caller probably said something that sounded like a new question, or was silent. The idle hangup (see Idle & Timeout above) closes the call after the check-in goes unanswered.

A refused action is never spoken. If the system declines a tool action during a call, the caller does not hear an error line. The bot simply answers.

---

Troubleshooting: Transfer Not Working

Bot says "transfer is not available" even though a transfer number is configured:

1. Check that Enable IVR Bridge Call is active on your subscription (IVR Settings → Voice Configuration → Escalation section)

2. Verify a Fallback Transfer Number is saved in IVR Settings for that phone line. Without one the bot cannot transfer the call at all. (CLI: velaro ivr line-update <id> --transfer-number +1..., MCP: ivr_update_line.)

3. If both are set and it still doesn't work, the subscription flag may need to be enabled — contact support

---

Pronunciation Overrides (sound rules for saying and hearing)

Use a pronunciation rule when the voice bot says a word wrong (brand name, acronym, part number) or hears a word wrong (the caller says one thing and the transcript shows another).

Go to Channels → IVR → Pronunciation → Add and fill in:

• Written form(s): every spelling that should trigger the rule, comma separated (e.g., TriTool, Tri-Tool, Tri Tool). Joined-capital names also match their spaced and hyphenated forms unless Exact match is on.
• Voice, how to say it: what the bot speaks (e.g., Try Tool, Fiz-ick, A P I). Spoken output only; transcripts and chat text keep the original word.
• Transcript text: what the AI and the transcript get when the caller says it (e.g., "fizzicks" becomes "Physics"). Caller side only.
• Team: optional; blank means every team.

Fill in at least one of Voice or Transcript text. Site rules beat the shared defaults, and a change reaches live calls within about 15 minutes. Emails and web addresses are already spoken as words ("name at company dot com"). When a customer says the bot mispronounces or mishears something, offer this rule first, and point them to the help article "Fix How the Phone Bot Says and Hears a Word".