customers
All customer stories
Top Voice AI companies are building with Assembly.
resources
Latest Release
Universal 3.5 Pro
Our new flagship speech-to-text model handles real-world audio the way it actually happens. Available for real-time and pre-recorded audio.
resources
The most consequential decision in building a voice agent isn't which LLM you use or what the agent sounds like — it's the architecture. An honest framework for three paths: a fully managed API, a bring-your-own-stack streaming pipeline, and the Sync API for turn-based audio. All three lead to AssemblyAI.
Three paths
One account, one SDK, one bill across all three.
Paths
3
Managed, bring-your-own-stack, and turn-based — all on the same accuracy foundation.
Bring your own
$0.45/hr
Universal-3.5 Pro Realtime STT — transcription only, unlimited concurrency.
Source: Pricing
Never re-platform
1
One API key, SDK, and billing relationship across every path.
Customer story · Report 4
“We're excited to make AssemblyAI's Universal-3.5 Pro available on LiveKit Inference. What really stands out is their pace of innovation with Context Carryover — it intelligently applies conversation context to improve transcription accuracy in a way most speech models don't, removing the need for users to predefine key terms.”
— David Zhao, Co-Founder, LiveKit
The decision
Audio comes in, intelligence happens, audio goes out. The real questions are how many pieces you want to own, and how audio reaches the model — a live stream, or finished turns you send yourself. Most teams over-weight "easy" or "flexible" and under-weight the tradeoffs that actually bite: time-to-production, per-layer control, cost structure, debugging surface, and scaling.
Recommended
Managed — one WebSocket
$4.50 /hr
STT + LLM + TTS included
Bring your own stack
Streaming STT — your LLM and TTS
$0.45 /hr
Transcription only
Turn-based
Short audio — one HTTP request
Short clips
Runs on Universal-3.5 Pro
Decision framework
The right architecture depends on how much of the stack you want to own and how audio reaches the model. Compare the dimensions that decide it below.
| Dimension | Voice Agent API | Universal-3.5 Pro Realtime | Sync API |
|---|---|---|---|
| How audio arrives | Live stream (WebSocket) | Live stream (WebSocket) | Finished turns (HTTP POST) |
| You own | Nothing — managed | LLM + TTS + orchestration | Turn detection + LLM + TTS |
| Setup | Agent in an afternoon | Swap STT in minutes | One POST |
| Pricing | $4.50/hr all-in | $0.45/hr + your LLM/TTS | Usage-based |
| Session resume | 30s built in | Via your orchestrator | Stateless (n/a) |
| Best for | Shipping fast | Existing stacks, scale | Dictation, IVR, own VAD |
All three share the same Universal-3.5 Pro Realtime accuracy foundation — the choice is about how much of the stack you own, not quality.
Migration
The paths use WebSocket or HTTP and standard JSON. Teams often prototype on the Voice Agent API for speed, then move individual components to Universal-3.5 Pro Realtime as they want more control — or reach for the Sync API where request/response fits better than a socket.
Switching means updating endpoints and message handling, not rebuilding. One account, one SDK, one bill. AssemblyAI is infrastructure, not a platform that captures you.
Typical path
Same account, same SDK — move pieces when you're ready.
Ship on the Voice Agent API today and move pieces to your own stack whenever you're ready.
Three, all on AssemblyAI's Universal-3.5 Pro Realtime foundation: a fully managed Voice Agent API (one WebSocket for STT + LLM + TTS), a bring-your-own-stack cascading pipeline with Universal-3.5 Pro Realtime as the STT layer (you bring the LLM and TTS), and — for turn-based short audio — the Sync API, a single HTTP request that returns a finished transcript.
Use the managed Voice Agent API if you want the fastest path to production — one API, one bill, a working agent in an afternoon. Build your own cascading pipeline with Universal-3.5 Pro Realtime if you already run an orchestrator (LiveKit, Pipecat, Vapi), need per-layer control over the LLM and TTS, or are optimizing margin at high scale. Both share the same accuracy foundation, so it's a control-vs-speed decision, not a quality one.
A cascading (or "chained") architecture wires separate STT, LLM, and TTS services together through an orchestrator: speech-to-text transcribes the caller, an LLM reasons over the transcript, and text-to-speech voices the reply. Universal-3.5 Pro Realtime is the STT layer in this pattern, working natively with LiveKit, Pipecat, Vapi, or any custom WebSocket orchestrator.
Use the Sync API when you do your own turn detection and just need a finished transcript per utterance — dictation, IVR menus, call routing, push-to-talk, or serverless functions that can't hold a socket open. It's a single HTTP POST with no WebSocket or session to manage. For a continuous live conversation, use the Voice Agent API or Universal-3.5 Pro Realtime streaming instead.
Yes — that's by design. The paths use WebSocket or HTTP with standard JSON, so switching means updating connection endpoints and message handling, not rebuilding. Teams often prototype on the Voice Agent API for speed, then move components to Universal-3.5 Pro Realtime as they want more control. One account, one SDK, one bill across all paths.
The Voice Agent API is a flat $4.50/hr with STT, LLM, and TTS included. Universal-3.5 Pro Realtime is $0.45/hr for transcription only (you add your own LLM and TTS), with add-ons billed as used. The Sync API is usage-based on Universal-3.5 Pro. See the pricing page for details.
