§6

Voice agent orchestration and realtime infrastructure

A speech model list stops where the hard part starts. Shipping a voice agent means solving voice activity detection, semantic endpointing, barge-in, sub-second round trips, WebRTC or WebSocket transport, SIP trunking and PSTN bridging, plus reconnects, jitter and concurrency — none of which an STT or TTS API gives you.

Data checked 2026-09-06

The orchestration layer is the thing that owns the audio loop between the caller and your LLM, and by 2026 it is a crowded market with three distinct shapes: open-source frameworks you run (LiveKit Agents, Pipecat), hosted platforms that rent you the whole loop per minute (Vapi, Retell, Bland, Ultravox), and telephony carriers or speech vendors extending upward into orchestration (Twilio ConversationRelay, Telnyx, SignalWire, Deepgram, ElevenLabs). The per-minute economics differ by more than 3x across these, and so does how much of the latency budget you actually control. This chapter also records two casualties, because the category churns fast enough that a stale shortlist wastes a sprint.

A How to choose

Start with the question of who owns the agent loop. If your business logic is a few prompts, a knowledge base and some webhooks, a hosted platform (Vapi, Retell, Bland, Telnyx) gets you to a live phone number in an afternoon and you will never regret it; if voice is the product and you will spend the next year tuning endpointing thresholds, interruption recovery and reconnect behavior, you want LiveKit Agents or Pipecat, because on a hosted platform you are renting somebody else's latency budget and cannot instrument inside it. Second axis: is this a phone product or an app product?

Twilio ConversationRelay, Telnyx and Bland are phone-shaped and give you nothing for an in-browser or in-app agent; LiveKit and Agora are WebRTC-shaped and treat SIP as a bridge; Pipecat is the only one that is genuinely transport-agnostic. Third axis: bundled versus unbundled per-minute pricing. Bland at ~$0.12/min all-in and SignalWire at $0.16/min are simple and expensive; LiveKit Cloud at $0.01/min agent time plus at-cost inference is cheap and requires you to model four line items — run your real minute volume through both before committing, because the crossover is usually somewhere between 50k and 200k minutes/month.

Fourth axis: speech-to-speech versus cascaded STT→LLM→TTS. Speech-native stacks (Hume EVI, Ultravox, Deepgram's Advanced tier, the realtime APIs underneath several of these) win on latency and prosody but cost you deterministic control — if you need strict script compliance, auditable tool calls and swappable voices, stay cascaded. Now the anti-recommendations.

Do not self-host LiveKit's SFU, Redis and SIP bridge just to place outbound calls; buy the carrier stack. Do not choose Vapi or Retell if you have a hard sub-500ms target you must prove in production, because you cannot profile their internals. Do not pick ElevenLabs Agents or Cartesia Line on orchestration merits — pick them when their voices are the reason you are here, and accept the tighter platform.

Do not put an enterprise CCaaS platform like Parloa in a shortlist next to a Python framework; it is a procurement decision with a different buyer and no public pricing. And verify liveness before you design around anything: Layercode's docs now redirect to an unrelated product and Vocode has not shipped a commit since 2024, both of which still rank well in search.

B At a glance

Name	Deployment	Transport & telephony	Model stack	Build interface	Concurrency	Pricing
LiveKit Agents	Self-hosted (Apache-2.0) or LiveKit Cloud	WebRTC native; SIP/PSTN via LiveKit SIP	BYO STT/LLM/TTS via plugins, or LiveKit Inference	Python and Node SDKs; Agent Builder (no-code)	5 free / 20 Ship / up to 600 Scale; unlimited self-hosted	Framework free (Apache-2.0). LiveKit Cloud: Build $0/mo (1,000 agent-session min, 5 concurrent, 5,000 WebRTC min); Ship $50/mo (5,000 min then $0.01/min, 20 concurrent); Scale $500/mo (50,000 min, up to 600 concurrent). WebRTC overage $0.0004–$0.0005/min; US local inbound $0.01/min; toll-free numbers $2.00/mo. Inference metered separately: STT $0.0025–$0.0117/min, TTS free–$0.03/min, LLM $0.0002–$0.0379/min.
Pipecat	Self-hosted anywhere; Pipecat Cloud for managed	Daily/SmallWebRTC, WebSocket, Twilio/Telnyx/Plivo PSTN, SIP	Fully BYO; 200+ provider integrations	Python SDK; JS/React/Android/iOS/React Native client SDKs	Unlimited (bounded by your own infrastructure)	Free (BSD-2-Clause). You pay only the vendors you plug in (transport, STT, LLM, TTS) and your own compute.
Pipecat Cloud	Managed cloud only (agents remain portable Pipecat code)	WebRTC, WebSocket, integrated PSTN and SIP	Fully BYO; optional bundled billing across 80+ providers	Python (Pipecat) plus a deploy CLI	Unlimited	Agent compute per active minute: agent-1x (0.5 vCPU/1GB) $0.01, agent-2x (1 vCPU/2GB) $0.02, agent-3x (1.5 vCPU/3GB) $0.03; reserved capacity ~50% off. Transport: 1:1 WebRTC voice free; voice+video $0.004/participant-min; SIP dial-in/out $0.003–$0.02/min; PSTN $0.018/min; SIP REFER transfer $0.20/event. Krisp VIVA free to 10k min/mo then $0.0015/min. Recording $0.005/min audio, $0.01349/min audio+video, storage $0.003/min.
Vapi	Managed cloud only	PSTN via Twilio/Telnyx/Vonage or Vapi numbers; SIP; web SDK	BYO vendors, passed through at cost (free with your own keys)	REST API, server webhooks, dashboard, web/mobile client SDKs	10 lines included, $10/line/mo after	$0.05/min platform fee for calls; $0.005 per SMS/chat message. Model, STT and TTS costs passed through at cost, or $0 to Vapi if you bring your own API keys. Concurrency: 10 lines included, $10/line/mo beyond. Add-ons: HIPAA $2,000/mo, Zero Data Retention $1,000/mo. Scale plan is a negotiated fixed platform fee against committed volume.
Retell AI	Managed cloud; dedicated servers on enterprise only	PSTN and SIP trunking (Retell or BYO carrier); web calls	Choice of LLMs and voices, each metered per minute	REST API, custom LLM webhook, dashboard flow builder	20 free, then $8 per concurrent call/mo; uncapped on enterprise	$0.07–$0.31/min all-in depending on LLM. Components: voice infrastructure $0.055/min, Retell platform TTS $0.015/min, US telephony $0.015/min. LLM examples: GPT-4.1 $0.045/min, Claude Sonnet $0.08/min, Gemini Flash $0.081/min, GPT-5.5 $0.16/min. Add-ons: knowledge base +$0.005/min, advanced denoising +$0.005/min, PII removal +$0.01/min, branded call +$0.10/outbound call. Subscriptions: concurrency $8/concurrent call/mo above 20 free, knowledge base $8/mo above 10 free, phone number $2/mo, verified number $10/mo. Chat agents from $0.002/message. $10 free credit.
Bland AI	Managed cloud; on-prem/VPC on enterprise	PSTN inbound/outbound, warm transfer; phone-centric	Bundled and proprietary; no BYO model or voice	Conversational Pathways graph builder plus REST API	10 / 50 / 100 by tier; sized to volume on enterprise	Start $0.14/min with $0 platform fee (10 concurrent); Build $0.12/min + $299/mo (50 concurrent); Scale $0.11/min + $499/mo (100 concurrent); Enterprise custom. Transfer minutes $0.05/$0.04/$0.03 by tier. All tiers include LLM, STT and TTS with no token charges. Free trial: 2 credits plus an inbound number (stated $15/mo value), no card required.
Deepgram Voice Agent API	Managed API, dedicated single-tenant, in-VPC, or self-hosted	WebSocket; native telephony integration for inbound/outbound	Deepgram STT + TTS bundled; BYO LLM and BYO TTS at lower rates	WebSocket API with JSON config and function calling	unknown (plan-dependent; enterprise negotiated)	Billed on WebSocket connection time. Standard: $0.056/min pay-as-you-go and $0.051/min on Growth through 2026-09-12, rising to $0.075 and $0.068/min after. Standard BYO-LLM: $0.050/$0.041 rising to $0.065/$0.059. Standard BYO-TTS: $0.065/$0.051. Advanced tier: $0.122/$0.110 rising to $0.163/$0.146. $200 free credit for new accounts; enterprise custom.
ElevenLabs Agents	Managed cloud only	WebSocket/WebRTC web widget and SDKs; Twilio/SIP trunk for phone	ElevenLabs TTS + STT bundled; BYO LLM billed at cost	Dashboard agent config, REST API, web/mobile SDKs, tools/webhooks	4–40 by tier; burst above ceiling at $0.16/min	$0.080/min standard, $0.160/min burst above your tier's concurrency; text messages $0.003 each. Tiers: Free (15 min, 4 concurrent), Starter $6/mo (75 min, 6), Creator $22/mo (275 min, 10), Pro $99/mo (1,238 min, 20), Scale $299/mo (3,738 min, 30), Business $990/mo (12,375 min, 40). LLM and telephony billed separately at cost.
Twilio ConversationRelay	Managed Twilio service; your WebSocket app is self-hosted	Twilio Voice PSTN/SIP in; wss:// to your server	STT Deepgram (default)/Google; TTS ElevenLabs (default)/Google/Amazon; BYO LLM	TwiML noun plus a WebSocket message protocol	Bounded by your Twilio account limits and your own service	ConversationRelay $0.07/min, on top of Twilio Voice: US outbound local/toll-free $0.0140/min, browser/app $0.0040/min, inbound local $0.0085/min (+$1.15/mo per number), inbound toll-free $0.0220/min (+$2.15/mo). LLM costs are yours. Volume discounts apply automatically at scale.
Telnyx Voice AI Agents	Managed cloud on Telnyx's own network	Native PSTN and SIP trunking; carrier-owned	Bundled STT/TTS (Deepgram, Telnyx, Rime, Inworld, Murf, Resemble); LLM metered separately	REST API, portal assistant builder, webhooks/tools	unknown (account- and plan-dependent)	$0.05/min voice engine (orchestration + STT + TTS). LLM tokens extra at Telnyx GPU rates — open models ~$0.004/min; Anthropic/OpenAI at their standard rates. Telephony from $0.0032/min US local inbound, outbound SIP trunking from $0.005/min. Telnyx quotes a realistic production agent near $0.056/min all-in. Plans: pay-as-you-go with no minimum; Committed $500/mo minimum with volume discounts; Enterprise $5,000/mo minimum. HIPAA included on all plans; SOC 2, GDPR, PCI DSS, ISO 27701.
Cartesia Line	Open-source SDK, self-run or deployed to Cartesia cloud via CLI	WebSocket/WebRTC; Cartesia-provisioned phone numbers	Cartesia Sonic TTS and Ink STT; 100+ LLMs via LiteLLM	Python SDK plus deploy CLI	8 concurrent on Free; higher on paid tiers	Voice agents $0.06/min of call duration on all paid plans, plus $0.014/min when using a Cartesia-provided phone number. Plans: Free $0/mo (1 provisioned number, 8 concurrent calls), Pro $5/mo with $5 prepaid agent minutes, Startup $49/mo with $49 prepaid, Scale $299/mo with $299 prepaid, Enterprise custom.
Ultravox Realtime	Managed cloud; model weights open but platform closed	WebSocket/WebRTC calls; SIP at $0.005/min	Ultravox speech-native LLM (bundled), not a swappable cascade	REST API, client SDKs, tools and call stages	5 on pay-as-you-go; no caps on Pro	Pay-as-you-go: 30 free minutes/month, then $0.05/min; 5 concurrent calls. Pro $100/mo: no concurrency caps, 5 custom voices, outbound call scheduler, SIP at $0.0048/min (vs $0.005/min PAYG). Threads billed at $2.00 per 1M input tokens (cached input free) and $15.00 per 1M output tokens. Enterprise custom with priority SLA.
Agora Conversational AI Engine	Managed cloud on Agora's global RTC network	WebRTC only; no first-class SIP/PSTN	BYO LLM/STT/TTS from a supported menu; some integrations beta	REST API to join an AI agent to an Agora channel; RTC SDKs	unknown (account-dependent)	$0.10/min for audio conversational AI tasks. Agora states this rate includes usage of the selected ASR, LLM and TTS models and applies even when you supply your own API keys. RTC minutes bill separately under standard Agora pricing; no free minute allowance is published for the engine itself.
SignalWire AI Agent	Managed cloud on SignalWire's own carrier and media stack	Native PSTN, SIP and WebRTC; owns carrier relationships	Bundled STT/LLM/TTS inside the AI runtime	SWML documents plus SWAIG tool handlers (Python type hints)	unknown (no per-seat licenses or stated minimums)	$0.16/min for the AI runtime, which SignalWire states includes STT, LLM inference, TTS, orchestration, state management and governance. Transport billed separately at carrier rates. No setup fees, minimums or per-seat licenses.
Hume EVI	Managed cloud only	WebSocket only; no native WebRTC or SIP	Hume EVI 3 / EVI 4-mini speech-to-speech; optional partner LLM	WebSocket API with TypeScript/Python SDKs and configs	Tier-dependent; higher limits on Business and Enterprise	Per-minute EVI usage beyond included minutes: Free $0.06/min (5 min included), Starter $0.07/min, Creator $0.07/min, Pro $0.06/min (1,200 min included), Scale $0.05/min, Business $0.04/min, Enterprise custom. Supplemental partner LLM costs are additional.
Parloa	Managed enterprise SaaS	Contact-centre telephony and CCaaS integrations; voice and chat	Managed by Parloa; not developer-selectable	Agent design, simulation and monitoring studio	Enterprise-scale; contract-defined	unknown — no pricing is published; sales-led with custom contracts.
Vocode	Self-hosted library only; hosted platform retired	Twilio/Vonage phone, WebSocket, Zoom (as of 2024)	BYO STT/LLM/TTS, pinned to 2024-era provider SDKs	Python library	Bounded by your own infrastructure	Library is MIT-licensed and free. Hosted platform pricing no longer published; the product site redirects to GitHub.
Layercode (now Toyo AI)	Was managed cloud; product URLs now redirect elsewhere	unknown (previously WebSocket/WebRTC with telephony)	unknown	JS/React client SDKs, last updated February 2026	unknown	unknown — the pricing page redirects to an unrelated product.

C Entries

LiveKit Agents

An agent worker process joins a LiveKit room as a participant, so the same code serves a browser, a mobile app or a phone call bridged in over SIP. Version 1.8 ships a trained turn-detector model rather than raw silence-based VAD, plus interruption handling, and plugins for most STT/LLM/TTS vendors; LiveKit Inference lets you call models through LiveKit Cloud without holding separate API keys. The distinction from Pipecat is that LiveKit gives you the media server too — it is one vertically integrated stack from SFU to agent — where Pipecat is a pipeline library that expects you to bring a transport.

Deployment	Self-hosted (Apache-2.0) or LiveKit Cloud
Transport & telephony	WebRTC native; SIP/PSTN via LiveKit SIP
Model stack	BYO STT/LLM/TTS via plugins, or LiveKit Inference
Build interface	Python and Node SDKs; Agent Builder (no-code)
Concurrency	5 free / 20 Ship / up to 600 Scale; unlimited self-hosted

Watch out: Self-hosting means running an SFU, Redis, agent workers and a SIP bridge — real infrastructure work that most teams underestimate. The Node SDK consistently trails the Python one in features. Cloud billing has both a monthly floor and four separate metered line items (agent minutes, WebRTC minutes, telephony minutes, inference), which makes cost modelling harder than a single all-in per-minute rate.

Framework free (Apache-2.0). LiveKit Cloud: Build $0/mo (1,000 agent-session min, 5 concurrent, 5,000 WebRTC min); Ship $50/mo (5,000 min then $0.01/min, 20 concurrent); Scale $500/mo (50,000 min, up to 600 concurrent). WebRTC overage $0.0004–$0.0005/min; US local inbound $0.01/min; toll-free numbers $2.00/mo. Inference metered separately: STT $0.0025–$0.0117/min, TTS free–$0.03/min, LLM $0.0002–$0.0379/min. · open source

Pipecat

Pipecat models a conversation as frames flowing through a pipeline of processors — transport in, VAD, STT, context aggregation, LLM, TTS, transport out — so you can intercept, reorder or replace any stage, including the interruption path. It ships a smart turn-detection model, Pipecat Flows for scripted conversation graphs, and transports covering Daily WebRTC, a dependency-free SmallWebRTC, raw WebSockets and Twilio/Telnyx/Plivo telephony. Where LiveKit hands you an opinionated stack including the media server, Pipecat hands you the plumbing and lets you choose everything else; v1.8.1 shipped in August 2026 and the repo has ~15k stars with daily commits.

Deployment	Self-hosted anywhere; Pipecat Cloud for managed
Transport & telephony	Daily/SmallWebRTC, WebSocket, Twilio/Telnyx/Plivo PSTN, SIP
Model stack	Fully BYO; 200+ provider integrations
Build interface	Python SDK; JS/React/Android/iOS/React Native client SDKs
Concurrency	Unlimited (bounded by your own infrastructure)

Watch out: Python only on the agent side — there is no Node or Go runtime, only client SDKs for web, React Native, iOS and Android. You own deployment, autoscaling, cold starts and observability unless you buy Pipecat Cloud. The 200+ integrations vary a lot in maintenance quality, and the frame/processor abstraction takes real time to internalise before you can debug a latency regression.

Free (BSD-2-Clause). You pay only the vendors you plug in (transport, STT, LLM, TTS) and your own compute. · open source

Pipecat Cloud

Daily runs your containerised Pipecat agents, scales them and bills only for active session time in three vCPU/memory sizes, with integrated PSTN and SIP so you do not need a separate carrier account. Concurrency is uncapped, reserved capacity cuts the compute rate roughly in half, and Krisp VIVA noise elimination is free up to 10k session minutes a month. It is the cheapest managed compute in this chapter because it is only compute — unlike Vapi or Retell there is no bundled model markup and no agent-builder dashboard, which is exactly the trade if you already have Pipecat code.

Deployment	Managed cloud only (agents remain portable Pipecat code)
Transport & telephony	WebRTC, WebSocket, integrated PSTN and SIP
Model stack	Fully BYO; optional bundled billing across 80+ providers
Build interface	Python (Pipecat) plus a deploy CLI
Concurrency	Unlimited

Watch out: It only runs Pipecat agents, so it is not a general voice platform and it deepens your commitment to one framework. You still buy and pay every model vendor separately, so the headline $0.01/min is a fraction of the real per-minute cost. Compute is sized in fixed vCPU tiers, so a pipeline doing local VAD or turn detection needs the 2x or 3x tier and the cost gap to bundled platforms narrows.

Agent compute per active minute: agent-1x (0.5 vCPU/1GB) $0.01, agent-2x (1 vCPU/2GB) $0.02, agent-3x (1.5 vCPU/3GB) $0.03; reserved capacity ~50% off. Transport: 1:1 WebRTC voice free; voice+video $0.004/participant-min; SIP dial-in/out $0.003–$0.02/min; PSTN $0.018/min; SIP REFER transfer $0.20/event. Krisp VIVA free to 10k min/mo then $0.0015/min. Recording $0.005/min audio, $0.01349/min audio+video, storage $0.003/min.

Vapi

Vapi owns the realtime loop — VAD, endpointing, barge-in, call state — and exposes it as an assistant config plus webhooks, with a dashboard, squads for multi-agent handoff and workflow branching. You still choose the STT, LLM and TTS vendors and pay them at cost, or zero platform markup if you supply your own keys, so the $0.05/min is a clean orchestration fee rather than a bundled black box. Compared with Retell it is more developer-API-shaped and less contact-centre-shaped, and compared with LiveKit you give up all control of the audio path in exchange for shipping this week.

Deployment	Managed cloud only
Transport & telephony	PSTN via Twilio/Telnyx/Vonage or Vapi numbers; SIP; web SDK
Model stack	BYO vendors, passed through at cost (free with your own keys)
Build interface	REST API, server webhooks, dashboard, web/mobile client SDKs
Concurrency	10 lines included, $10/line/mo after

Watch out: HIPAA and zero-data-retention are $2,000/mo and $1,000/mo add-ons, which is a large step for a small healthcare pilot. Concurrency is metered at $10 per line per month, so bursty inbound traffic gets expensive in a way per-minute pricing hides. There is no self-hosted option, and when you hit a latency or turn-taking regression you are limited to Vapi's own call logs — you cannot instrument inside the pipeline.

$0.05/min platform fee for calls; $0.005 per SMS/chat message. Model, STT and TTS costs passed through at cost, or $0 to Vapi if you bring your own API keys. Concurrency: 10 lines included, $10/line/mo beyond. Add-ons: HIPAA $2,000/mo, Zero Data Retention $1,000/mo. Scale plan is a negotiated fixed platform fee against committed volume.

Retell AI

Retell publishes its stack as separate meters — $0.055/min for the voice infrastructure, $0.015/min for platform TTS, $0.015/min US telephony, plus whichever LLM you pick — so a working agent lands somewhere between $0.07 and $0.31 per minute depending on model choice. The platform side is aimed at operations teams: knowledge bases, post-call analysis, batch outbound campaigns, verified/branded caller ID, PII removal and advanced denoising, each metered as its own add-on. It overlaps heavily with Vapi; the practical difference is that Retell leans further into contact-centre features and per-feature metering, where Vapi leans into a single orchestration fee and developer webhooks.

Deployment	Managed cloud; dedicated servers on enterprise only
Transport & telephony	PSTN and SIP trunking (Retell or BYO carrier); web calls
Model stack	Choice of LLMs and voices, each metered per minute
Build interface	REST API, custom LLM webhook, dashboard flow builder
Concurrency	20 free, then $8 per concurrent call/mo; uncapped on enterprise

Watch out: The unbundled meters add up fast and every useful feature — knowledge base, denoising, PII redaction — is a separate per-minute surcharge, so quoted floor prices are misleading. Self-hosting and dedicated infrastructure are enterprise-only. As with any hosted platform you cannot see or tune the endpointing and interruption logic, which matters if your callers are on noisy mobile connections.

$0.07–$0.31/min all-in depending on LLM. Components: voice infrastructure $0.055/min, Retell platform TTS $0.015/min, US telephony $0.015/min. LLM examples: GPT-4.1 $0.045/min, Claude Sonnet $0.08/min, Gemini Flash $0.081/min, GPT-5.5 $0.16/min. Add-ons: knowledge base +$0.005/min, advanced denoising +$0.005/min, PII removal +$0.01/min, branded call +$0.10/outbound call. Subscriptions: concurrency $8/concurrent call/mo above 20 free, knowledge base $8/mo above 10 free, phone number $2/mo, verified number $10/mo. Chat agents from $0.002/message. $10 free credit.

Bland AI

Bland runs its own vertically integrated stack rather than brokering third-party models, which is why it can quote one all-in per-minute number with no token line items. Agents are built as Conversational Pathways — a graph of nodes with conditions and transfers — rather than as code, and warm transfers to humans are billed at a separate lower transfer-minute rate. It is the simplest cost model in this chapter and the least flexible: you cannot swap in a specific TTS voice vendor or LLM the way you can on Vapi, Retell or LiveKit.

Deployment	Managed cloud; on-prem/VPC on enterprise
Transport & telephony	PSTN inbound/outbound, warm transfer; phone-centric
Model stack	Bundled and proprietary; no BYO model or voice
Build interface	Conversational Pathways graph builder plus REST API
Concurrency	10 / 50 / 100 by tier; sized to volume on enterprise

Watch out: You cannot bring your own models or voices — the vertical integration that produces the flat rate also removes vendor choice, and the underlying model stack is not disclosed. The product is built around phone calls, so it is a poor fit for embedded in-app or browser voice. Mid tiers carry $299–$499/mo platform fees before a single minute, and on-prem/VPC is enterprise-only.

Start $0.14/min with $0 platform fee (10 concurrent); Build $0.12/min + $299/mo (50 concurrent); Scale $0.11/min + $499/mo (100 concurrent); Enterprise custom. Transfer minutes $0.05/$0.04/$0.03 by tier. All tiers include LLM, STT and TTS with no token charges. Free trial: 2 credits plus an inbound number (stated $15/mo value), no card required.

Deepgram Voice Agent API

Instead of a framework, Deepgram sells the loop itself: you open a single WebSocket, stream audio in, and Deepgram runs its STT, calls an LLM, synthesises with Aura/Flux and streams audio back, handling barge-in, endpointing and turn-taking prediction inside the connection. Function calling lets your server run tools mid-conversation, and you can substitute your own LLM or TTS provider at a discounted rate while keeping Deepgram's orchestration. Deployment options run from the managed API to single-tenant, in-VPC and fully self-hosted, which is unusual in this category and the main reason regulated buyers look here.

Deployment	Managed API, dedicated single-tenant, in-VPC, or self-hosted
Transport & telephony	WebSocket; native telephony integration for inbound/outbound
Model stack	Deepgram STT + TTS bundled; BYO LLM and BYO TTS at lower rates
Build interface	WebSocket API with JSON config and function calling
Concurrency	unknown (plan-dependent; enterprise negotiated)

Watch out: There is no agent builder, campaign tooling or CRM surface — everything above the audio loop is your code, so it is not a fast path for non-engineers. A published price increase lands on 2026-09-12, roughly a third above current rates, so budget on the post-increase numbers. The economics assume Deepgram STT; going fully BYO on both LLM and TTS erodes most of the reason to be here.

Billed on WebSocket connection time. Standard: $0.056/min pay-as-you-go and $0.051/min on Growth through 2026-09-12, rising to $0.075 and $0.068/min after. Standard BYO-LLM: $0.050/$0.041 rising to $0.065/$0.059. Standard BYO-TTS: $0.065/$0.051. Advanced tier: $0.122/$0.110 rising to $0.163/$0.146. $200 free credit for new accounts; enterprise custom.

ElevenLabs Agents

Formerly marketed as Conversational AI, this is ElevenLabs wrapping its TTS with turn-taking, interruption handling, tools, knowledge bases and web/phone transports so you do not build the loop yourself. Pricing is subscription-shaped rather than pure usage: each tier includes a block of minutes and a concurrency ceiling, and exceeding the ceiling triggers burst billing at double the standard rate. Choose it when the specific ElevenLabs voice is a product requirement; on pure orchestration merits it is more constrained and, at $0.08/min before LLM and telephony, more expensive than assembling the same thing on Pipecat or LiveKit.

Deployment	Managed cloud only
Transport & telephony	WebSocket/WebRTC web widget and SDKs; Twilio/SIP trunk for phone
Model stack	ElevenLabs TTS + STT bundled; BYO LLM billed at cost
Build interface	Dashboard agent config, REST API, web/mobile SDKs, tools/webhooks
Concurrency	4–40 by tier; burst above ceiling at $0.16/min

Watch out: Concurrency is tied to subscription tier and tops out at 40 on the published Business plan, with overflow charged at 2x — which punishes exactly the bursty inbound traffic voice products get. LLM and telephony are extra on top of $0.08/min, so the real cost is meaningfully higher than the headline. Orchestration is less configurable than a framework, and the whole thing assumes you want ElevenLabs synthesis.

$0.080/min standard, $0.160/min burst above your tier's concurrency; text messages $0.003 each. Tiers: Free (15 min, 4 concurrent), Starter $6/mo (75 min, 6), Creator $22/mo (275 min, 10), Pro $99/mo (1,238 min, 20), Scale $299/mo (3,738 min, 30), Business $990/mo (12,375 min, 40). LLM and telephony billed separately at cost.

Twilio ConversationRelay

A `<ConversationRelay>` noun inside `<Connect>` hands a live call to Twilio's realtime service, which transcribes the caller, sends structured messages to your WebSocket, synthesises whatever text you send back, and manages interruption (the `interruptible` attribute) and DTMF. Twilio supplies the speech layer from a fixed menu — Deepgram or Google for STT, ElevenLabs, Google or Amazon for TTS — while your server owns the LLM, memory and tools entirely. It is now GA and it is the lowest-commitment way to add an LLM to an existing Twilio voice application, because none of your numbers, routing or compliance setup changes.

Deployment	Managed Twilio service; your WebSocket app is self-hosted
Transport & telephony	Twilio Voice PSTN/SIP in; wss:// to your server
Model stack	STT Deepgram (default)/Google; TTS ElevenLabs (default)/Google/Amazon; BYO LLM
Build interface	TwiML noun plus a WebSocket message protocol
Concurrency	Bounded by your Twilio account limits and your own service

Watch out: It is not an agent runtime — you must host and scale a WebSocket service, and own conversation state, tools and the LLM loop yourself, so it saves you the audio plumbing and nothing else. STT and TTS are limited to Twilio's provider menu, and it is phone-only: there is no browser or in-app agent story here. Access requires onboarding and accepting Twilio's Predictive and Generative AI addendum.

ConversationRelay $0.07/min, on top of Twilio Voice: US outbound local/toll-free $0.0140/min, browser/app $0.0040/min, inbound local $0.0085/min (+$1.15/mo per number), inbound toll-free $0.0220/min (+$2.15/mo). LLM costs are yours. Volume discounts apply automatically at scale.

Telnyx Voice AI Agents

Telnyx owns its own network and GPU inference, so the agent, the speech models and the carrier minutes all bill on one invoice and the media never leaves their stack — which is the latency and cost argument. The $0.05/min voice engine covers turn-taking, interruption handling, tool calls and knowledge-base retrieval plus STT (Deepgram or Telnyx) and a wide TTS menu; LLM tokens are billed separately at Telnyx GPU rates, and open models land near $0.004/min. HIPAA is included on every plan rather than sold as an add-on, which is a real differentiator against Vapi's $2,000/mo.

Deployment	Managed cloud on Telnyx's own network
Transport & telephony	Native PSTN and SIP trunking; carrier-owned
Model stack	Bundled STT/TTS (Deepgram, Telnyx, Rime, Inworld, Murf, Resemble); LLM metered separately
Build interface	REST API, portal assistant builder, webhooks/tools
Concurrency	unknown (account- and plan-dependent)

Watch out: The economics depend on using Telnyx's own models and network; bringing a frontier LLM erases much of the price advantage. Committed and Enterprise tiers impose $500/mo and $5,000/mo minimums. The developer community and third-party ecosystem are far smaller than LiveKit's or Pipecat's, and there is little here for a browser or in-app WebRTC agent.

$0.05/min voice engine (orchestration + STT + TTS). LLM tokens extra at Telnyx GPU rates — open models ~$0.004/min; Anthropic/OpenAI at their standard rates. Telephony from $0.0032/min US local inbound, outbound SIP trunking from $0.005/min. Telnyx quotes a realistic production agent near $0.056/min all-in. Plans: pay-as-you-go with no minimum; Committed $500/mo minimum with volume discounts; Enterprise $5,000/mo minimum. HIPAA included on all plans; SOC 2, GDPR, PCI DSS, ISO 27701.

Cartesia Line

Line is Cartesia's Apache-2.0 Python SDK for wrapping a text agent in voice: it handles audio orchestration, real-time interruption and turn-taking, tool calling, multi-agent handoff and built-in web search, then deploys with a CLI to Cartesia's managed runtime with observability. LiteLLM underneath means 100+ LLM providers work, but the speech layer is Cartesia's own Sonic TTS and Ink STT — this exists to make their models easy to build agents on. It is the newest and smallest of the open frameworks here (roughly 100 GitHub stars against Pipecat's 15k), so treat the ecosystem accordingly.

Deployment	Open-source SDK, self-run or deployed to Cartesia cloud via CLI
Transport & telephony	WebSocket/WebRTC; Cartesia-provisioned phone numbers
Model stack	Cartesia Sonic TTS and Ink STT; 100+ LLMs via LiteLLM
Build interface	Python SDK plus deploy CLI
Concurrency	8 concurrent on Free; higher on paid tiers

Watch out: Young and small — around 100 stars and a fraction of the integrations, examples and community answers you get with Pipecat or LiveKit, which matters when you hit an edge case at 2am. Managed deployment is Cartesia-only, and the speech layer is Cartesia's, so it is a poor fit if you want to A/B different TTS vendors. Python-only, and the free plan's 8 concurrent calls is a hobby ceiling.

Voice agents $0.06/min of call duration on all paid plans, plus $0.014/min when using a Cartesia-provided phone number. Plans: Free $0/mo (1 provisioned number, 8 concurrent calls), Pro $5/mo with $5 prepaid agent minutes, Startup $49/mo with $49 prepaid, Scale $299/mo with $299 prepaid, Enterprise custom. · open source

Ultravox Realtime

Ultravox is a speech-native model — audio goes straight into the LLM without a separate transcription step — wrapped in a hosted agent platform with calls, tools, stages and an outbound scheduler. Skipping the STT hop is the latency argument, and the flat $0.05/min with $0.005/min SIP is among the cheapest bundled rates in this chapter. The underlying model weights are published under MIT, though that repo has been quiet since December 2025; the platform itself is closed and hosted.

Deployment	Managed cloud; model weights open but platform closed
Transport & telephony	WebSocket/WebRTC calls; SIP at $0.005/min
Model stack	Ultravox speech-native LLM (bundled), not a swappable cascade
Build interface	REST API, client SDKs, tools and call stages
Concurrency	5 on pay-as-you-go; no caps on Pro

Watch out: Because the model is speech-native, you cannot freely swap TTS voices or drop in a different LLM the way a cascaded pipeline lets you — you take Ultravox's speech behaviour or you leave. The open-weight model repository has had no pushes since December 2025, which is worth checking before you treat the open weights as an exit path. No self-hosting of the platform is offered, and the free tier caps at 5 concurrent calls.

Pay-as-you-go: 30 free minutes/month, then $0.05/min; 5 concurrent calls. Pro $100/mo: no concurrency caps, 5 custom voices, outbound call scheduler, SIP at $0.0048/min (vs $0.005/min PAYG). Threads billed at $2.00 per 1M input tokens (cached input free) and $15.00 per 1M output tokens. Enterprise custom with priority SLA.

Agora Conversational AI Engine

Agora bolts an agent orchestrator onto its established global RTC network, joining an AI participant to an Agora channel and handling interruption with an acoustic algorithm tuned for real network conditions. You bring the models — OpenAI, Azure OpenAI, Gemini, Vertex, Claude, Dify or a custom LLM; ARES, Azure or Deepgram for STT; Azure, ElevenLabs, Cartesia, OpenAI or Hume for TTS. The reason to choose it over LiveKit is coverage: if your users are on unreliable mobile networks in Asia, South America or Africa, Agora's edge network and last-mile handling are a genuinely different product from a generic SFU.

Deployment	Managed cloud on Agora's global RTC network
Transport & telephony	WebRTC only; no first-class SIP/PSTN
Model stack	BYO LLM/STT/TTS from a supported menu; some integrations beta
Build interface	REST API to join an AI agent to an Agora channel; RTC SDKs
Concurrency	unknown (account-dependent)

Watch out: At $0.10/min this is the most expensive flat orchestration fee here, and it applies even when you bring your own model keys — you pay Agora's rate on top of whatever you owe your model vendors. Several TTS integrations are still marked beta. It is WebRTC-first with no meaningful SIP/PSTN telephony story, and the tooling assumes you are building against Agora's RTC SDKs rather than writing Python agent code.

$0.10/min for audio conversational AI tasks. Agora states this rate includes usage of the selected ASR, LLM and TTS models and applies even when you supply your own API keys. RTC minutes bill separately under standard Agora pricing; no free minute allowance is published for the engine itself.

SignalWire AI Agent

SignalWire's argument is architectural: the AI kernel executes inside its media stack rather than as an orchestration layer relaying HTTP between separate vendors, and it owns its own carrier relationships for PSTN, SIP and WebRTC. Agents are defined in SWML with tool calling through SWAIG, where you declare tools with Python type hints and your backend handler enforces the rules before results reach the model. SWML is deliberately TwiML-compatible enough that existing Twilio flows can migrate, which is the main reason a Twilio shop would look here.

Deployment	Managed cloud on SignalWire's own carrier and media stack
Transport & telephony	Native PSTN, SIP and WebRTC; owns carrier relationships
Model stack	Bundled STT/LLM/TTS inside the AI runtime
Build interface	SWML documents plus SWAIG tool handlers (Python type hints)
Concurrency	unknown (no per-seat licenses or stated minimums)

Watch out: SignalWire's own published figure is ~1,200ms average conversational turn latency (1.09–1.46s), which is well above the sub-500ms target teams in this category chase — check that against your use case before believing the in-media-stack latency argument. At $0.16/min the AI runtime is the priciest bundled rate in this chapter. You author in SWML and SWAIG rather than a general-purpose Python agent framework, so portability off the platform is poor.

$0.16/min for the AI runtime, which SignalWire states includes STT, LLM inference, TTS, orchestration, state management and governance. Transport billed separately at carrier rates. No setup fees, minimums or per-seat licenses.

Hume EVI

EVI measures tone, rhythm and timbre in the caller's voice and conditions its own delivery on them, returning expression measures alongside the transcript so your application can react to how something was said. EVI 3 is an integrated English speech-language model; EVI 4-mini covers 11 languages but requires a supplemental LLM, and both can be paired with partner models from Anthropic, OpenAI, Google or Fireworks. Turn-taking and interruption are handled inside the WebSocket session, capped at 30 minutes per session — it competes on emotional expressiveness, not on telephony or orchestration breadth.

Deployment	Managed cloud only
Transport & telephony	WebSocket only; no native WebRTC or SIP
Model stack	Hume EVI 3 / EVI 4-mini speech-to-speech; optional partner LLM
Build interface	WebSocket API with TypeScript/Python SDKs and configs
Concurrency	Tier-dependent; higher limits on Business and Enterprise

Watch out: Sessions are hard-capped at 30 minutes, which rules out long-running assistants without reconnect logic. It is WebSocket-only with no native WebRTC or SIP, so any telephony deployment means you build and operate the bridge yourself. EVI 3 is English-only, and if you do not specifically need prosody awareness, a cascaded stack does the same job with more control and usually less cost.

Per-minute EVI usage beyond included minutes: Free $0.06/min (5 min included), Starter $0.07/min, Creator $0.07/min, Pro $0.06/min (1,200 min included), Scale $0.05/min, Business $0.04/min, Enterprise custom. Supplemental partner LLM costs are additional.

Parloa

Parloa is an AI Agent Management Platform aimed at contact centres rather than developers: the product covers the whole lifecycle — designing agents, simulating and testing them at scale, deploying across voice and chat, and monitoring and optimising them in production. It leans on multilingual handling and real-time translation, and carries the certifications enterprise procurement asks for (ISO 27001:2022, SOC 2 Type 1 and 2, PCI DSS, HIPAA, DORA), with an SAP endorsement and named references including TUI. It belongs in this chapter as the enterprise CX end of the market — the buyer is a contact-centre owner replacing IVR, not an engineer adding voice to a product.

Deployment	Managed enterprise SaaS
Transport & telephony	Contact-centre telephony and CCaaS integrations; voice and chat
Model stack	Managed by Parloa; not developer-selectable
Build interface	Agent design, simulation and monitoring studio
Concurrency	Enterprise-scale; contract-defined

Watch out: No public pricing and a sales-led procurement cycle, so it cannot be trialled the way any other entry here can. You get no low-level access to VAD thresholds, endpointing or the audio pipeline — the abstraction is a managed agent, not a framework. It is heavy overkill if you need one agent inside a product, and the platform assumes contact-centre surroundings (CRM, CCaaS, workforce tooling) you may not have.

unknown — no pricing is published; sales-led with custom contracts.

Vocode

Vocode was the first widely adopted open-source library for streaming LLM conversations onto phone calls and Zoom, and at 3.8k stars it still ranks highly in search results for voice agent frameworks. It is no longer maintained: the last push to vocode-core was 2024-11-15 and the last release, v0.1.114a2, was August 2024. The vocode.dev and www.vocode.dev domains now 301-redirect to the GitHub organisation, which indicates the hosted platform is no longer being sold. Listed here so nobody loses a sprint adopting it — the live equivalents are Pipecat and LiveKit Agents.

Deployment	Self-hosted library only; hosted platform retired
Transport & telephony	Twilio/Vonage phone, WebSocket, Zoom (as of 2024)
Model stack	BYO STT/LLM/TTS, pinned to 2024-era provider SDKs
Build interface	Python library
Concurrency	Bounded by your own infrastructure

Watch out: Unmaintained for roughly two years, so it pins old provider SDK versions and will not have integrations for any model released since late 2024, nor any of the semantic turn-detection work that defines the current generation. No security patching, no community support, and the commercial entity behind it appears to have wound the product down. Do not start new work on it.

Library is MIT-licensed and free. Hosted platform pricing no longer published; the product site redirects to GitHub. · open source · discontinued

Layercode (now Toyo AI)

Layercode sold a hosted voice pipeline where your own backend supplied the agent logic over a webhook while Layercode handled transport, turn-taking and speech. As of September 2026 both layercode.com and docs.layercode.com return a 301 to toyo.ai, which sells an AI executive assistant for founders — a different product in a different category — and the GitHub organisation layercodedev now displays the name Toyo. The client SDKs (layercode-js-sdk, layercode-react-sdk) last saw updates in February 2026. Treat the voice infrastructure product as gone.

Deployment	Was managed cloud; product URLs now redirect elsewhere
Transport & telephony	unknown (previously WebSocket/WebRTC with telephony)
Model stack	unknown
Build interface	JS/React client SDKs, last updated February 2026
Concurrency	unknown

Watch out: Not a viable choice: the product and documentation URLs redirect to an unrelated assistant product, and the client SDKs stopped receiving updates in early 2026. Confidence is medium rather than high because no formal shutdown notice was published — the conclusion is inferred from domain redirects, the GitHub organisation rename and SDK inactivity. If you have a live Layercode integration, contact the vendor directly before assuming continuity.

unknown — the pricing page redirects to an unrelated product. · discontinued