The voice is the first thing your client hears and the first thing their customer judges. Operators obsess over the prompt and the knowledge base — both matter — and then ship a default TTS voice that sounds like a 2019 GPS. The voice is not a cosmetic choice. It's a latency decision, a trust decision, and a brand decision, all at once. Here's how to pick one on purpose.
The three things a voice trades off
Every voice is a point on a triangle, and you can't max all three:
- Latency — how fast the first audio comes back. Under ~1 second and the call feels human; north of that and callers start talking over the agent. Some ultra-natural voices are slow to synthesize; some fast voices sound flat.
- Naturalness — prosody, breaths, emotion, how it handles a comma. This is what makes a caller forget they're talking to software.
- Stability — does it mispronounce the client's product names, choke on numbers, or randomly change tone mid-sentence? A voice that's stunning 90% of the time and bizarre 10% of the time is worse than a plain voice that's consistent.
Pick the two that matter most for the use-case and accept the third. A high-volume inbound support line lives or dies on latency and stability. A luxury med-spa concierge can spend a little latency budget on naturalness.
The providers, honestly
- ElevenLabs — the naturalness benchmark. Emotional range and pronunciation are best-in-class, and the voice library is huge. The tradeoff is latency and cost; their newer low-latency models (Flash-tier) close most of the gap and are what you'd actually run in production on Vapi or Retell.
- Cartesia (Sonic) — the latency play. Purpose-built for real-time voice agents, sub-100ms-class synthesis, genuinely natural. This is increasingly the default when the call cannot tolerate lag. First stop for high-volume phone work.
- PlayHT / Play — solid middle ground, good low-latency options, widely integrated.
- Platform-native voices — Vapi, Bland, Retell, and Synthflow all ship built-in voices (including OpenAI and Deepgram TTS). Fine for a demo or a cost-sensitive build; you'll usually upgrade to ElevenLabs or Cartesia for anything client-facing where the brand matters.
Whatever the platform, the voice is a swappable setting. Don't marry the default.
How to actually choose
- Match the brand, not your taste. A personal-injury law firm wants calm, competent, reassuring. A pizza chain wants upbeat and quick. A dermatology clinic wants warm and unhurried. Write down the three adjectives the client's brand should project, then audition voices against those, not against which one you personally like.
- Gender and age should fit the audience, not defaults. Test it — sometimes the "obvious" choice underperforms with the actual caller demographic.
- Audition on the real script, over a real phone call. A voice that sounds gorgeous in the browser preview can fall apart compressed through the telephony codec (8kHz, narrowband). Always listen to it on an actual phone call through your platform before you decide. This step catches more problems than any spec sheet.
- Stress-test the hard words. Feed it the client's product names, drug names, street names, dollar amounts, phone numbers, and email addresses. This is where voices embarrass you. If it says the company name wrong, you'll need pronunciation overrides (phoneme tags or an alias dictionary) — build those before launch, not after the client complains.
The settings that quietly matter
- Speaking rate — most defaults are slightly too fast for a phone call. Nudging it down a notch improves comprehension more than swapping the voice.
- Stability vs. expressiveness sliders (ElevenLabs-style) — for support and scheduling, bias toward stability; a predictable voice beats a theatrical one that occasionally goes weird.
- Filler words and backchannels — a well-placed "mm-hm" or "let me check that" makes the agent feel present, but overdone it's grating. Tune it to the use-case.
The honest caveat
The voice is a real lever but it is not a fix for a bad agent. A perfect voice on top of a slow, guessing, poorly-scoped agent still loses the call — the caller just gets frustrated in a nicer tone. Get latency, knowledge scope, and escalation right first, then spend real time choosing a voice that fits the brand and survives a phone call. Pick it on purpose, audition it on the phone, and stress-test the hard words. That's the whole discipline.