Building Production Voice AI Agents: Latency, Architecture, and What Nobody Tells You

Total320–560ms

Latency Budget by Layer

Voice Activity Detection (10–30ms)

VAD decides when the user has stopped speaking and the pipeline should fire. A misconfigured VAD is the single easiest way to add 500ms of latency without touching any model. Most implementations default to a trailing silence window of 500–800ms — that pause sits entirely in the user experience before a single API call fires.

In production, configure VAD with:

• Silence threshold: 300ms for call center contexts, 200ms for high-tempo applications
• Endpointing: fire on silence, not on a fixed timer
• Echo cancellation: required whenever the agent speaks; browser getUserMedia handles this with echoCancellation: true

Deepgram's streaming STT includes built-in VAD endpointing via endpointing=300 — use this rather than a separate VAD layer, as it eliminates an additional round-trip.

STT: Streaming Transcription (60–120ms)

Batch transcription — send audio, wait for full transcript — adds 600–1,200ms before your LLM call even starts. This alone makes sub-250ms unreachable. The solution is streaming STT with interim results.

Deepgram Nova-3 (released 2025) delivers streaming transcription with a first-word latency around 60–80ms over WebSocket — meaningfully faster than Nova-2's 80–120ms, with improved accuracy on accented speech and noisy environments. You do not wait for the complete transcript; you begin processing on is_final: true utterances:

User audio → WebSocket → Deepgram Nova-3 (streaming)
                              ↓
                    interim results (ignored)
                              ↓
                    is_final: true → LLM pipeline fires

Critical configuration: punctuate=true, smart_format=true, and endpointing=300. Without endpointing set, Deepgram uses server-side silence detection that defaults longer than your VAD window. For lowest latency, use model=nova-3 with tier=nova explicitly set.

LLM Reasoning (150–250ms)

LLM first-token latency is the hardest constraint to optimize. GPT-4 in streaming mode cannot reliably hit sub-200ms first-token in typical network conditions. The model choices that achieve 100–200ms in practice in 2026:

• GPT-4o-mini — ~120–150ms first-token median; the default choice for most voice turn completions
• Claude Haiku 4.5 — ~100–150ms first-token; strongest instruction-following for structured voice turns; handles healthcare and fintech prompts cleanly
• Gemini 2.0 Flash — ~100–140ms first-token; competitive on throughput, useful as a fallback
• Groq-hosted Llama 3.3 70B — sub-80ms first-token via custom inference hardware; the choice when raw latency is the constraint; model quality adequate for most voice use cases
• GPT-4o — ~200–300ms first-token; reserve for complex reasoning turns where quality matters more than speed

Stream the response. Pass tokens to TTS as they arrive — do not buffer the full LLM output before starting TTS synthesis. The overlap between LLM generation and TTS synthesis recovers 100–200ms of total latency.

For prompt engineering: keep system prompts under 400 tokens for voice, strip all markdown formatting (it degrades TTS output), and keep total context under 2,000 tokens where possible — token count has a near-linear relationship with first-token latency.

TTS: Streaming Synthesis (40–100ms)

Two providers dominate production voice AI in 2026 for latency-sensitive workloads:

Cartesia Sonic delivers first-audio-chunk in 40–60ms — the fastest production TTS available. It uses a diffusion-based architecture that generates audio differently from autoregressive models, which is why the latency floor is lower. The trade-off: voice cloning fidelity is somewhat behind ElevenLabs. For voice AI agents where naturalness matters but clone quality is secondary, Sonic is the right call.

ElevenLabs Flash (eleven_flash_v2_5) delivers first-audio-chunk in 60–100ms with higher voice quality and the most realistic cloning available. The configuration that matters for latency:

• Model: eleven_flash_v2_5 — not the standard model, which runs 200–400ms
• Streaming: stream=true
• Output format: pcm_16000 for telephony, mp3_44100_128 for browser
• Streaming latency optimization: optimize_streaming_latency=4 (aggressive mode)

Use streaming TTS: do not wait for the complete audio file before playback. The client begins playing as soon as the first audio chunk arrives. For browser clients, the Web Audio API handles chunked playback natively; for telephony, use RTP packetization.

Network Transport (20–60ms)

With a well-configured WebRTC connection, transport adds 20–40ms round-trip. With a WebSocket-only approach through a distant cloud region, transport alone can add 200ms in the tail. This is where the transport choice has the most impact.

Full Stack Architecture

The production architecture for a sub-300ms voice AI agent:

The agent worker sits between the media plane and the model APIs. It receives raw audio frames from LiveKit, streams them to Deepgram, fires the LLM on final utterances, and pushes TTS audio frames back into the LiveKit room. The client never calls model APIs directly — this is essential for PII control and rate-limit management.

ICE Trickle and the LiveKit SFU Pattern

Why ICE Trickle Matters

WebRTC connection establishment uses Interactive Connectivity Establishment (ICE) to find a network path between peers. In the naive implementation — wait for all ICE candidates before signaling — setup latency adds 500–2,000ms to every call start. This is invisible in demos and very visible in production.

ICE Trickle solves this: candidates are sent to the remote peer as they are gathered, and connectivity checks begin immediately. Call setup time drops to 100–400ms in most network conditions.

LiveKit implements ICE Trickle automatically. What you need to deploy:

• STUN servers — used for reflexive candidate discovery; stun.l.google.com:19302 works for most cases; deploy your own for HIPAA environments to keep traffic off third-party infrastructure
• TURN servers — required for clients behind symmetric NAT, common in enterprise networks; LiveKit's hosted tier includes TURN, or deploy coturn yourself
• Signaling — LiveKit's built-in signaling server handles offer/answer exchange; no separate WebSocket signaling server required

LiveKit SFU Pattern

A Selective Forwarding Unit receives encoded media streams and forwards them to participants without decoding and re-encoding. For voice AI, this matters because:

1. The agent worker receives RTP packets from the SFU rather than raw WebRTC — simpler to handle in server-side Python or Node.js code
2. Multiple agents or observers can subscribe to the same audio stream without additional encoding cost
3. The SFU handles DTLS/SRTP complexity; the agent sees plain RTP internally

The LiveKit room model maps cleanly to a voice call session:

from livekit import agents, rtc
import asyncio

async def entrypoint(ctx: agents.JobContext):
    await ctx.connect()

    async for event in ctx.room.on("track_subscribed"):
        if event.track.kind == rtc.TrackKind.KIND_AUDIO:
            audio_stream = rtc.AudioStream(event.track)
            asyncio.create_task(process_audio(audio_stream, ctx.room))

async def process_audio(stream: rtc.AudioStream, room: rtc.Room):
    async for frame in stream:
        await pipeline.push_frame(frame)

LiveKit's agent framework handles room lifecycle, track subscription, and RTP framing. Application code focuses on pipeline logic.

WebRTC vs SIP: Which Transport to Use

This is the question that trips up most teams evaluating voice AI infrastructure. They are not competing choices — they solve different integration problems.

Use WebRTC when you control the client — a web app, mobile app, or embedded SDK. It gives you wideband Opus audio (meaningfully better STT accuracy), lower setup latency, and direct control over the media path.

Use SIP when the caller is on a real phone number — inbound calls to a support line, outbound dialer campaigns, or integration with an existing contact center (Genesys, Five9, Twilio PSTN). Twilio's Media Streams provides a WebSocket bridge from PSTN to your agent worker, which avoids running a full SIP stack yourself.

The G.711 codec limitation of PSTN calls has an underappreciated consequence: STT accuracy on 8kHz narrowband audio is meaningfully lower than on 16kHz+ wideband. For healthcare or fintech agents where transcription accuracy directly affects outcomes, browser/mobile WebRTC with Opus gives a material accuracy advantage over telephone calls.

A production voice AI WebRTC architecture typically uses both: WebRTC for app callers and a SIP trunk or Twilio Media Streams for inbound phone calls, with the same agent worker behind both paths.

Observability: What to Instrument

Voice AI pipelines fail silently. A WebRTC ICE failure looks like a dropped call. A Deepgram WebSocket disconnect looks like the agent not hearing the user. A TTS timeout manifests as silence on the line. Without structured observability, every incident is a multi-hour debugging session across three services.

Instrument the following at minimum:

Per-call latency histogram — record wall-clock time from VAD endpoint event to first TTS audio chunk, broken down by component: stt_latency_ms, llm_first_token_ms, tts_first_chunk_ms. Alert on p95 > 800ms for any single component.

Per-call transcription confidence — Deepgram returns a confidence score per utterance. Log confidence distributions; a degradation in median confidence correlates with audio quality issues, codec mismatches, or background noise problems before callers start complaining.

WebRTC ICE connection state — log ICE state transitions (checking → connected → disconnected → failed). Track failed rates by client region. Elevated failure rates in a specific geography usually indicate TURN server coverage gaps.

STT WebSocket reconnections — Deepgram WebSocket connections drop under load or network events. Count reconnections per call. A call with 3+ reconnections will have visible transcription gaps; flag and review these separately.

LLM error rates — log 4xx/5xx rates from your LLM provider independently from total call failure. A 429 spike during peak hours needs a different response (add capacity, queue calls) than a 500 (inspect payloads, contact provider).

Use structured logging with a call_id field on every log event. Voice AI incidents always span Deepgram, your agent worker, and your SFU. Without a consistent call_id, joining those log lines across services is impossible.

For a broader framework on instrumenting LLM workloads in production — what to log, how to set up Langfuse, and which metrics actually correlate with model quality regressions — see our LLMOps guide.

Building This for Production

The full stack — streaming STT, low-latency LLM, streaming TTS, WebRTC via LiveKit SFU, structured observability, and pre-LLM PII redaction — is the minimum viable architecture for a voice AI agent that holds up in production. The demo that skips any of these layers will surface its gaps within the first 200 calls.

The latency numbers are achievable. Sub-400ms p50 is not theoretical — it is what Prodinit operates for a healthcare scheduling platform handling 2000+ calls per day in a HIPAA-covered environment. The architecture works because every layer is optimized independently and composed correctly.