Choose models and voices

Every turn of a call goes through three vendors you pick: one hears the caller, one decides the reply, one speaks it. The choice is the single biggest lever on how a call feels — more than the prompt — because it sets the latency between the caller finishing a sentence and the agent starting one. Above about a second that gap reads as hesitation; above two, callers start talking over the agent.

The live catalogue with prices is on Models and voices. This page is about choosing.

What each provider decides

ProviderDecidesLatency it addsCost driver
STTWhether the caller’s words arrive correctly — names, numbers, accents, noisesmall; streams while they talkper minute
LLMWhat the agent says and does; how well it follows the promptthe largest share — time to first tokenper token, every turn
TTSHow the agent sounds; whether it’s pleasant to listen to for five minutestime to first audioper character

A mistake at STT propagates: if “fourteenth” arrives as “fortieth”, the best LLM books the wrong date. A slow LLM can’t be fixed by a fast TTS. Choose in that order.

Choosing an LLM

The trade is speed against judgment. For phone work, speed wins more often than people expect: a fast model that replies in 400 ms with a decent answer beats a brilliant one that pauses for two seconds.

  • Start with a small, fast model from any vendor — the ones the catalogue describes as low-latency or “for voice”. Most flows (reminders, booking, triage, surveys) never need more.
  • Move up only for a specific reason: multi-step reasoning inside a turn, long function results the model must interpret, or strict policy compliance where a wrong answer is costly.
  • Reasoning models think before replying. That’s a strength for analysis and a liability for live conversation — the thinking time lands in the pause before every reply. Prefer non-reasoning models for the call itself.

Settings on the LLM node

SettingWhat it doesGuidance
ModelWhich modelSee above
Temperature (0–2)Randomness of wording0.3–0.7 for a natural but consistent agent. Some models don’t accept a temperature; Talkif hides or strips it for those — setting one has no effect
Max TokensCap on reply lengthKeep it low (a few hundred). It’s a safety net, not a style control — use the prompt to ask for short replies
Priority Tier (OpenAI)Faster, more reliable inference at 2× the token priceWorth it for customer-facing lines under load; not for testing

Choosing STT

The trade is language coverage and robustness against price.

  • Language first. Pick a model that lists the caller’s language; multilingual, auto-detecting models (Soniox, xAI) are the safe default for mixed lines.
  • Numbers and names. The newer conversational models (Deepgram nova-3, for example) handle spelled names and digit strings noticeably better — relevant for reference numbers, postcodes, dates of birth. Test with your own data.
  • Telephony audio is 8 kHz. Every model in the catalogue works with it; the ones that advertise telephony support (xAI, Amazon Transcribe) were tuned on it.

Settings on the STT nodes — each takes a Model and, where the model supports several, a Language (Deepgram, Google, Amazon Transcribe, Cartesia; a language hint for ElevenLabs). Soniox and xAI models detect language automatically and expose only the model choice.

Choosing TTS

The trade is naturalness against latency, and it’s the most subjective choice — always listen before you decide.

  • Try two or three voices with the same prompt in the Flow Tester. The same words land very differently.
  • Flash / low-latency models for conversation; the expressive premium models are for lines where tone carries the brand and a little extra delay is acceptable.
  • Match language and accent to the callers, not to the company.

Settings on the TTS nodes

ProviderSettings
ElevenLabsVoice; Stability 0–1 (higher = steadier, lower = more expressive); Similarity 0–1; Style 0–1; Speaker Boost; Speed 0.7–1.2. On v3 models only Stability applies, and the prompt can use inline delivery tags — see Write prompts
CartesiaVoice; Emotion (on sonic-3); Speed 0.6–1.5; Volume 0.5–2.0
RimeVoice; Emotion; Speed; Volume
xAIVoice; Speed 0.7–1.5
GoogleVoice only

A default that works

If you have no constraints yet: Deepgram nova-3-general → a small, fast OpenAI model (the templates use gpt-4o-mini) at temperature 0.5 → ElevenLabs eleven_flash_v2_5 with a voice you’ve listened to. That’s what the built-in templates ship with, and it’s a good baseline to measure others against: change one provider at a time, run the same test conversation, compare the per-turn latency in the Flow Tester and the cost line on the call.

Provider and model choices are part of the flow definition, so they’re versioned with it. Switching the LLM and publishing creates a new version; a number pinned to the old version keeps the old model until you republish over it. See Flows and versions.

Next