Voice quality issues — robotic AI, dropped audio, latency

Voice quality issues

The AI receptionist should sound natural — calm pace, expressive tone, conversational rhythm. When it sounds robotic, choppy, or unusually slow, the cause is one of three things: ElevenLabs model selection, network conditions on the caller's side, or a misconfigured voice profile. Here's how to diagnose each.

Symptom 1 — The AI sounds flat or robotic

The voice should sound like a real person. If it sounds clearly synthetic — flat affect, weird pacing, mispronounced names — the most common cause is the voice model in your settings.

Open *Settings → Voice → Voice profile*. Three things to check:

1. Model: ElevenLabs offers eleven_turbo_v2.5 (fast, slightly less expressive) and eleven_multilingual_v2 (slightly slower, much more expressive). For most US customers, multilingual_v2 is the better choice — the latency hit is ~200ms but the difference in naturalness is night and day. Switch and call yourself to compare.
2. Voice: The voice ID you've selected. Some voices in the catalog are explicitly "broadcaster" or "narrator" voices that sound polished but stilted in conversation. The voices labeled "Conversational" are tuned for back-and-forth dialogue and almost always sound better for phone calls. Three we recommend by default: Bri (warm, mid-range, neutral US accent), Marcus (calm, lower register), Lila (energetic, slightly higher register).
3. Stability slider: The default 0.5 is right for most cases. Drop to 0.35 for more expressiveness (better for casual brands), raise to 0.65 for more consistency (better for medical/legal where you want predictable enunciation).

Symptom 2 — The AI cuts off mid-sentence or skips words

This is almost always a network problem on the caller's side. The AI streams audio in 200ms chunks; if the caller's connection drops packets, individual chunks arrive late or not at all and the audio sounds choppy.

You can confirm this from the dashboard: open the call in *Activity*, expand the Network panel. It shows packet loss and jitter for the call. >2% packet loss or >50ms jitter explains the symptom.

If multiple recent calls show high packet loss, the issue is on your business's side, not the caller's — usually a saturated office Wi-Fi or a congested ISP link. Run an internet speed test from the office at peak hours; if upload <5 Mbps or jitter >30ms, voice will be unreliable across many calls.

Symptom 3 — There's a long pause before the AI responds

A tiny pause (200–400ms) is normal — that's the AI listening for end-of-utterance. A long pause (>1.5s) usually means one of:

• Long inputs. If the caller speaks for 30 seconds straight, the AI waits for a clear pause before responding. Encourage natural turn-taking by greeting the caller with a short opener, not a 5-sentence menu.
• First-call cold start. The very first call to a freshly-deployed voice agent takes ~800ms longer to start because the model warms up. Subsequent calls are fast. If every call is slow, this isn't the issue.
• Hand-off retrieval. If the AI is searching your knowledge base for a specific answer, that lookup adds 100–400ms. If it consistently adds >1s, your KB is unusually large or your *Settings → AI → Search behavior* is set to "exhaustive" — switch to "fast" for sub-second response.

Symptom 4 — The AI mispronounces my business name or industry terms

Add pronunciation hints in *Settings → Voice → Pronunciation*. You give the AI a phonetic spelling for any term it gets wrong; the next call onwards it uses the corrected pronunciation.

Examples:

• Business "Schaeffer" → "SHAY-fer"
• Industry "Worcestershire" → "WOOSS-tər-shər"
• "Houston" (the Texas city, not the NYC street) → "HEW-stən"

You can add up to 50 pronunciation hints per agent. The hints don't expire and they apply to all future calls.

Symptom 5 — The AI sounds completely different on every call

The voice ID is being randomized somewhere in your config, which usually means an integration is overriding it on each call. Check *Settings → Integrations* — if you have a third-party tool that injects per-call voice settings, it may be sending an empty voice_id field, causing fallback to a random voice from the pool. Either fix the integration or unset it from the BizRnR side.

Still bad?

Email support@bizrnr.com with a specific call ID where the issue was clear. We can replay the call's audio internally, inspect the model parameters at the moment of call, and identify exactly what went wrong. Most voice complaints get resolved with a single setting tweak.