Every turn of a conversation travels through the same five stages. Each one streams into the next, nothing waits for a complete buffer before the next stage starts.
The call arrives over your telephony provider at 8 or 16kHz audio, depending on the trunk and codec.
Speech is transcribed as it is spoken, Sarvam saaras:v3 for Indian languages or Deepgram nova-3 for global ones. Endpointing decides when you've finished talking.
The model starts generating the reply token-by-token as soon as the transcript finalizes, instead of waiting to compose the whole answer.
Synthesis begins on the first LLM tokens, Sarvam bulbul:v2, ElevenLabs, or Cartesia, streaming audio chunks back as they're generated.
Audio chunks are streamed back down the WebRTC or SIP leg to the caller, without waiting for the full sentence to be synthesized.
Telephony failover runs across 8 providers: Exotel, Vobiz, Twilio, Plivo, Telnyx, Vonage, Cloudonix, Asterisk ARI. See the full provider matrix on AI Agent and Languages.
"Latency" in voice AI marketing usually means one of two very different measurements.
The gap between the LLM finishing its decision and the first audio frame reaching the caller. It's dominated by streaming chunking and TTS first-byte time, and it's the smaller of the two numbers because it ignores everything that happened before the model started answering.
What the caller actually feels: the moment they stop speaking to the moment the agent starts responding. Realistically ~500–800ms depending on STT finalization, LLM time-to-first-token, TTS time-to-first-byte, telephony round-trip, and deployment region.
You may see "<200ms" claims on voice AI marketing sites. That figure describes pipeline overhead at best, not end-to-end turn latency. If a vendor won't say which measurement they're quoting, or over which network and region it was measured, the number isn't comparable.
Turn latency isn't a property of a platform, it's a property of the exact stack and region you deploy. Six factors dominate it.
How aggressively the transcriber finalizes a turn. More conservative endpointing is more accurate on long pauses, but adds delay before the LLM ever sees text.
Time to first token varies by model, prompt size, tool-call depth, and provider load. Streaming output means the TTS doesn't wait for the full reply.
How quickly the synthesizer emits its first audio chunk. Streaming TTS pipelines start speaking mid-generation; chunked ones wait for sentence boundaries.
8kHz PSTN audio, codec conversion, and network distance between the caller's carrier and the pipeline region all add transport milliseconds.
A pipeline hosted closer to the caller's telephony POP cuts round-trip time on every stage of the turn, this is why region choice matters for India traffic.
Language-aware stacks (Sarvam for Indian languages, Deepgram/ElevenLabs for global) differ in how they endpoint, tokenize, and synthesize, and that changes the number.
The only latency figure worth trusting is the one measured on your calls, your region, and your provider stack. Crawlii's workspace dashboard logs per-tool latency and execution timings for every turn, so you can see exactly where each millisecond went instead of trusting a marketing table.
Run a live agent in the Playground, place a real call through one of the 8 telephony providers, and read the breakdown in your workspace, or walk through it with our team on a demo call.