Latency
Real, measured turn-latency numbers from Talkr's voice pipeline — STT, LLM, TTS, and total time from when a caller stops talking to when the agent responds.
These are real numbers pulled from Talkr's own call logs, not marketing estimates. The methodology and its limitations are below — read those before comparing against another platform's published numbers, since latency is highly sensitive to provider choice, geography, and network path.
What's measured
For every turn in a call, Talkr's pipeline (built on Pipecat) records:
- STT time-to-first-byte — from audio reaching the speech-to-text provider to the first transcript byte back.
- LLM time-to-first-token — from the transcript reaching the LLM to the first token of its response.
- TTS time-to-first-byte — from text reaching the voice provider to the first audio byte back.
- Total turn latency — from the moment the caller stops talking to the moment the agent's voice starts, the number that actually determines how a call feels.
This isn't a synthetic benchmark — it's the same instrumentation that runs on every real call, so the numbers below come from Talkr's own logged data.
Measured results
Small sample (4 test calls, 33 conversational turns) from a single test environment — not yet aggregated production call volume. Treat these as directionally accurate, not a statistically rigorous benchmark. This page will be refreshed as real call volume grows.
| Metric | Median (p50) | Average | Range |
|---|---|---|---|
| STT time-to-first-byte | 0.52s | 0.53s | 0.49s – 0.71s |
| LLM time-to-first-token | 1.01s | 1.06s | 0.77s – 1.63s |
| TTS time-to-first-byte | 0.38s | 0.38s | 0.37s – 0.40s |
| Total turn latency (plain conversation) | 2.22s | 2.28s | 2.02s – 2.83s |
| Total turn latency (turn includes a tool call) | 3.56s | 3.43s* | 2.98s – 3.87s* |
Why tool calls take longer: any turn where the agent calls a tool (looks up a knowledge base, hits an external API, checks availability) waits on that round-trip before it can respond — that's real, unavoidable latency on top of the base STT→LLM→TTS pipeline, not a Talkr overhead.
Methodology and caveats
Read these before drawing conclusions:
- Test calls, not production telephony. These numbers come from browser-based WebRTC test calls, not real phone calls over a telephony provider's PSTN/SIP network. WebRTC and telephony media paths have different network hops, jitter buffers, and codec transcoding — real inbound/outbound phone call latency will differ, and we'll publish that separately once there's enough real call volume to measure it honestly.
- Provider-specific. These calls ran on Talkr's own trial-credit managed AI stack: Deepgram for STT, Google Gemini 3.5 Flash for LLM, Deepgram for TTS. Talkr is BYOK — if you configure different providers or models, your numbers will differ. This page reflects one specific stack, not a platform-inherent ceiling or floor.
- Small sample. 4 calls is enough to sanity-check the instrumentation and give a directional sense of performance, not enough for a rigorous percentile claim (p95/p99 need far more data). We'll widen this as real call volume accumulates.
How this is measured internally
These are measurements, not estimates. Talkr times every stage of every turn as it happens — from the moment the caller stops speaking to the moment your agent starts — and records it against that call. Nothing is sampled, averaged after the fact, or inferred: what you see on a call is what that caller actually experienced.