Choosing a CPaaS for AI voice agents: STT, LLM, TTS, and telephony as one bill.
The defining change in CPaaS over the last 24 months is that the unit of work is no longer a phone call — it is a conversation conducted by software. An AI voice agent answers an inbound call (or places an outbound one), listens, decides what to say, says it, listens again, and repeats until the conversation ends. The underlying CPaaS still moves audio packets; what's new is the four-component stack sitting on top of those packets, and the question of which provider integrates which component into one billing relationship.
The four-component stack
Every production AI voice agent in 2026 is some configuration of the same four pieces:
- Telephony / media transport. The CPaaS that originates or terminates the call on the PSTN, streams audio in and out of the call leg, and handles DTMF, recording, and call control. This is the part CPaaS providers have always done.
- Speech-to-text (STT). The model that converts streaming audio into text the LLM can read. Deepgram, OpenAI Whisper, AWS Transcribe, and several CPaaS-native engines compete here.
- Large language model (LLM). The model that produces the agent's next utterance given conversation history. OpenAI, Anthropic, Google Gemini, and open-weight models running on inference platforms are all in production use.
- Text-to-speech (TTS). The model that converts the LLM's text back into audio. ElevenLabs, OpenAI TTS, Cartesia, PlayHT, and AWS Polly are the common choices.
End-to-end latency — the time from when the user stops speaking to when the agent starts — is dominated by LLM time-to-first-token, with STT, TTS, and network adding smaller fixed costs. Well-tuned production deployments in 2026 achieve 600 to 900 milliseconds of perceived latency. Below 700ms feels conversational. Above 1500ms feels broken.
Three integration models
CPaaS providers have converged on roughly three integration patterns. Picking among them is the central decision in voice-agent vendor selection.
Model 1: Media streaming + bring-your-own everything
The CPaaS streams audio in and out of the call over WebSockets or a similar bidirectional channel; you wire up STT, LLM, and TTS yourself, on whatever providers you prefer. This is the most flexible model and the most expensive in engineering time. Twilio Media Streams, SignalWire's bidirectional audio, and Plivo's audio streaming all fall here. You get to pick best-of-breed at each layer and you get four bills.
Model 2: Orchestrated voice agent platforms
The CPaaS provides a higher-level abstraction — typically called something like AI Studio, ConversationRelay, or Voice AI — that orchestrates STT, LLM, and TTS calls inside the call itself. You configure prompts and providers; the CPaaS handles plumbing. Twilio ConversationRelay, Vonage AI Studio, and several of Infobip's conversational products fit this model. Fewer bills, less control, and faster to ship.
Model 3: First-party AI inside the CPaaS
The CPaaS supplies its own STT, LLM, and TTS — at least as an option — alongside its telephony. Telnyx Inference and Telnyx Storage are the cleanest example: one account, one bill, the LLM and vector store sit next to the voice termination. The tradeoff is that the in-house models may not match best-of-breed providers on quality or feature surface. The economics, when they fit, are excellent.
What to evaluate
When shortlisting CPaaS providers for an AI voice-agent workload, the criteria diverge from a traditional voice CPaaS evaluation:
- Streaming audio latency. Round-trip from the CPaaS's media stream endpoint to your inference back to the call. Single-digit hundreds of milliseconds is achievable; older patterns can run higher.
- Barge-in handling. Whether the platform supports interrupting the agent mid-utterance when the user starts talking. Without barge-in, the agent feels mechanical.
- VAD quality. Voice activity detection — knowing when the user has finished speaking — is harder than it sounds and varies meaningfully across providers.
- Function calling / tool use. Whether the platform exposes a clean way for the LLM to call back into your application mid-conversation (book an appointment, look up an order, send a confirmation SMS).
- Recording and transcript handling. Compliance often requires call recording and full transcripts. Whether the platform supplies these natively or requires you to handle them affects integration cost.
- Billing consolidation. How many invoices you receive at the end of the month. Buyers underestimate this until they have to reconcile four vendors against one project.
Approximate cost build-up
A typical AI voice agent call in 2026 costs in the neighborhood of $0.06 to $0.18 per minute of conversation, depending on how the components are sourced. The build-up looks roughly like this:
| Component | Typical /min cost | Notes |
|---|---|---|
| CPaaS voice termination | $0.005 – $0.012 | Lower if carrier-owned (Telnyx, Bandwidth) |
| Speech-to-text (streaming) | $0.005 – $0.015 | Deepgram, Whisper, AWS Transcribe |
| LLM (assume 200 input + 100 output tokens / turn, 4 turns / min) | $0.020 – $0.080 | Wide range — GPT-class vs open-weight |
| Text-to-speech | $0.020 – $0.050 | ElevenLabs higher, AWS Polly lower |
| Carrier surcharges + recording + storage | $0.005 – $0.020 | Highly variable |
The LLM and TTS lines dominate. The carrier-grade voice savings that matter so much for high-volume non-AI workloads (Bandwidth vs Twilio, Telnyx vs Vonage) get diluted by the AI components. This shifts which providers are most cost-effective for AI voice agents relative to traditional voice workloads.
Build vs buy
Companies starting from scratch and shipping in weeks rather than months should default to a turnkey orchestrated platform — Twilio ConversationRelay, Vonage AI Studio, or a CPaaS-native AI like Telnyx. The engineering savings dwarf the per-minute premium for the first year of operation.
Companies running meaningful volume (tens of thousands of minutes per month and up) where per-minute economics start to matter, or who need fine-grained control over each component for product or compliance reasons, should plan to build on raw media streams. The engineering investment is real but pays back in unit economics and vendor optionality.
The voice-agent stack is the first time CPaaS providers have had to compete on AI quality, not just packet quality. Some are adapting fast. Others are extending their existing voice products with AI bolted on.
Where each major provider stands
A short tour of how the providers in this directory have positioned themselves for AI voice workloads, in alphabetical order:
Amazon Chime SDK
Strong fit for teams already in AWS — Transcribe, Bedrock-hosted LLMs, and Polly are all native AWS services that consolidate into one bill with Chime SDK voice. Less plug-and-play than dedicated voice-agent platforms; more flexible if you're comfortable in AWS.
Telnyx
Telnyx Inference (LLM serving) and Telnyx Storage (vector store) sit next to Telnyx Voice. The clearest example of CPaaS-native AI in one billing relationship. Carrier-grade voice economics help when the workload scales.
Twilio
The broadest ecosystem. ConversationRelay (managed agent orchestration) is the easiest path; Media Streams is the underlying primitive for bring-your-own. Voice Intelligence handles recording, transcription, and post-call analysis.
Vonage
Vonage AI Studio is a mature low-code voice-agent builder; the underlying APIs are competitive. The Ericsson parent positions Vonage interestingly as carrier network APIs (CAMARA) mature.
SignalWire
Strong fit for teams that want deep VoIP control. SWML can script complex agent behaviors declaratively. Smaller ecosystem of pre-built agent tooling than Twilio.
Plivo
Plivo Voice with audio streaming supports the standard bring-your-own pattern. Less marketing emphasis on AI as a category than some competitors, but the primitives are there.
Frequently asked
What is an AI voice agent?
An AI voice agent is software that conducts a spoken conversation with a human over a phone call or audio channel. The agent listens through speech-to-text, decides what to say using a large language model, speaks back through text-to-speech, and is connected to the PSTN by a CPaaS or telephony provider.
Which CPaaS providers support AI voice agents in 2026?
All major CPaaS providers now support AI voice agents in some form. Twilio offers ConversationRelay and Media Streams; Vonage has AI Studio; Telnyx offers a tighter bundle through Telnyx Voice plus Telnyx Inference and Storage; Plivo, SignalWire, Sinch, and Infobip all expose voice plus media-streaming APIs that can carry an AI agent.
What latency should I expect from an AI voice agent?
End-to-end latency from end-of-user-speech to start-of-agent-speech is the metric that matters. In 2026, well-tuned production deployments achieve 600 to 900 milliseconds. Below 700ms feels conversational. Above 1500ms feels broken. Latency is dominated by LLM time-to-first-token.
How is an AI voice agent billed?
Typically through three to four separate line items: a per-minute voice charge from the CPaaS, a per-minute speech-to-text charge, a per-token language-model charge, and a per-character or per-minute text-to-speech charge. Some providers consolidate this into fewer line items.
Should I build my own voice agent stack or use a turnkey provider?
Build it yourself if you need fine-grained control over each component, want vendor optionality, or are running at a scale where consolidated solutions become more expensive than the sum of best-in-class parts. Use a turnkey provider if you are starting from scratch, need to ship within weeks rather than months, or value one billing relationship over component optimization.
For full per-provider profiles, including starting pricing and capability tags, see the main directory. For pricing detail beyond AI voice — including A2P fees, carrier surcharges, and number rental — see the 2026 pricing report.