Is speech-to-speech faster than a cascade?

Not reliably. Hosted speech-to-speech models land at 0.6 to 1.5 s to first audio on the Artificial Analysis board; a tuned cascade lands at 0.6 to 0.9 s. See the speech-to-speech guide for the trade-offs.

What latency should I promise a customer?

Quote the p95 from your own call logs, on the phone network the customer will use. On Voiceflint that number is on every call page.

Guides · updated 9 September 2026

Voice agent latency: the 700 ms budget

People notice a pause of about a second in conversation and start talking over it at around 1.5 seconds. That gives a voice agent a budget of roughly 700 to 900 ms from the end of the caller's sentence to the first sound of the reply, and less on a phone line, which adds its own delay. This guide breaks that budget into the pieces you can actually change.

Where the milliseconds go

Stage	Typical	What moves it
End-of-turn decision	100–300 ms	Native end-of-turn STT (Flux) at the low end; silence timeout plus a turn model at the high end
Final transcript	0–200 ms	Zero with fused end of turn; otherwise the STT model's finalisation time
LLM time to first token	250–600 ms	Model size, prompt length, prompt caching, region, reasoning switched off
TTS time to first byte	40–250 ms	Flash models versus frontier voices
Network and audio	50–100 ms on WebRTC	Region of the worker relative to the caller
Phone network	+150–300 ms	Carrier and codec; not under your control

Voiceflint's presets carry a measured band: balanced at 700–900 ms, lowest-latency at 500–700 ms and highest-quality at 900–1200 ms voice to voice on WebRTC.

Language models by time to first token

Model	TTFT	Tokens / s	$ / min in a call
gpt-oss-120b on Cerebras	250 ms	3000	$0.0042
Gemini 2.5 Flash-Lite	290 ms	—	$0.0019
Llama 4 Scout on Groq	300 ms	600	$0.0021
GPT-4.1 nano	330 ms	—	$0.0010
GPT-5 nano	350 ms	—	$0.0005
Gemini 3.5 Flash-Lite	350 ms	356	$0.0063
Gemma 4 31B	350 ms	—	$0.0073
GPT-4.1 mini	380 ms	—	$0.0040
GPT-5 mini	400 ms	—	$0.0024
GPT-5.6 Luna	400 ms	113	$0.0023

Time to first token is the number that matters; tokens per second only matters once the first sentence is long. Reasoning models are excluded because their thinking time lands on top of every turn.

Measure it per turn, not per call

Averages hide the turns that hurt. Every Voiceflint call has a waterfall for each turn: end-of-turn, STT, LLM first token, TTS first byte and the total, so a slow p95 points at one stage rather than at the whole stack. If the LLM stage is the problem, try prompt caching or a smaller model before touching anything else.

Five things that reliably help

Fuse end-of-turn detection into STT. It is the largest single saving and costs nothing extra on Flux.
Keep the system prompt short and stable so the cached prefix hits every turn.
Ask the model for a short first sentence; the TTS can start on it while the rest streams.
Run the agent in the region closest to your callers. Voiceflint has US East, US West and EU West worker pools and routes provider traffic to regional endpoints where the provider offers them.
Turn off reasoning for conversational turns and reserve it for tool decisions if you need it at all.

Questions

Is speech-to-speech faster than a cascade?: Not reliably. Hosted speech-to-speech models land at 0.6 to 1.5 s to first audio on the Artificial Analysis board; a tuned cascade lands at 0.6 to 0.9 s. See the speech-to-speech guide for the trade-offs.
What latency should I promise a customer?: Quote the p95 from your own call logs, on the phone network the customer will use. On Voiceflint that number is on every call page.

Run the numbers on your own calls

500 free minutes a month, every model, a receipt on every call.

Start free Open the calculator