Physical Turnaround Latency Breakdown

Turn-by-turn ear-to-mouth timing. Shows where acoustic turnaround delay originates across the voice pipeline.

VAD / End-of-Turn Speech-to-Text LLM TTFT CRM Tool RTT TTS TTFB
Cloudflare Edge Native
Deepgram Flux · DeepSeek V4.1 Flash (1017) · Aura-2 · Colocated D1
390ms
LiveKit Agents
Silero VAD v5 · Deepgram Nova-3 · OpenAI gpt-6-luna · Inworld TTS-2 Flash
475ms
Retell AI Managed
Retell Endpointer · Deepgram Nova-3 · OpenAI gpt-6-luna · ElevenLabs Flash v2.5
510ms
Hover over or tap any pipeline segment to inspect component latencies and bottlenecks.
Fastest Turnaround 390ms Cloudflare Edge Native
Tool Execution 12ms vs 30ms Sub-15ms with colocated D1
Fastest TTFT 85ms Cloudflare (DeepSeek Flash)
Policy Safety 100% Pass 5/5 canonical guardrails

Unit Cost Economics & Scale Calculator

Unbundled component cost comparison and interactive volume modeling across platforms.

10,000 calls
5.0 min / call
Component Cloudflare Voice Agent LiveKit Cloud Inference Retell AI Managed
Platform & Orchestration $0.000 / min $0.004 / min (LiveKit Cloud) $0.050 / min (managed markup)
Speech-to-Text (STT) $0.007 / min (Deepgram Flux) $0.005 / min (Deepgram Nova-3) Included (Deepgram Nova-3)
LLM Intelligence $0.005 / min (DeepSeek Flash 1017) $0.038 / min (gpt-6-luna) $0.028 / min (gpt-6-luna)
Text-to-Speech (TTS) $0.010 / min (Deepgram Aura-2) $0.012 / min (Inworld TTS-2 Flash) $0.015 / min (ElevenLabs Flash v2.5)
Database & CRM Tool Compute $0.002 / min (D1 colocation) $0.002 / min (D1 Edge API) $0.002 / min (External Edge D1)
Effective Unit Cost $0.0240 / min $0.0610 / min $0.0950 / min
Calculated Monthly Spend $1,200 / mo $3,050 / mo $4,750 / mo
Annual Cloudflare Savings
$42,600 / year
Compared to Retell AI baseline at selected 10,000 calls / mo

Policy & Security Fidelity Scorecards

Verification of non-negotiable business rules across all 5 canonical regression scenarios.

Scope Guardrail: Pool Heater Refusal — Transcript Viewer

Select any scenario to inspect the turn-by-turn dialogue, policy guardrail assertions, and tool execution payloads.

Interactive WebRTC Multi-Agent Voice Console

Live microphone testing and ear-to-mouth verification across all three architectures. Evaluates acoustic turnaround, speech recognition, and real-time CRM tool invocation.

🔒 Security Protected

Demo Voice Engine — Security Gate

WebRTC audio calls stream bidirectionally to live AI backends (Retell AI, LiveKit, and Cloudflare Edge Durable Objects). To prevent unauthorized API consumption during public deployment, enter your developer security token or password to unlock audio calling.

Protected by DEMO_VOICE_TOKEN / SEED_TOKEN

Live CRM Tool Execution Benchmark

Directly measure mid-call tool execution latency across the Worker edge backend.

D1 Schedule Cache (Edge Native) Recommended
12ms (RTT 47ms)
Single-digit local D1 query execution with edge caching
Uncached Workiz CRM Path Legacy WAN
418ms (RTT 453ms)
Simulated Workiz multi-page API network delay
Inspect Raw Backend Response Payload
Ready to test. Click a measurement button above to dispatch a live CRM request.