Short answer: the 2026 AI-voice stack is no longer one purchase. Speech recognition (STT) is its own market, led by Deepgram and its Flux agent line. Synthesis (TTS) is its own market, with Cartesia's Sonic setting the low-latency bar. Orchestration is its own market too, split between open-source frameworks like Pipecat (maintained by Daily) and LiveKit Agents, and newer voice-native lanes such as Ultravox. And above all three sits the layer buyers actually sign for: the end-to-end specialists, the Vapi family of Vapi, Retell, Bland, and Synthflow, who sell the assembled agent as a single product. The decision that matters is not "which vendor" but "which layers do I own, which do I rent, and who holds the latency budget." This post is the decision-side view of that stack. It is not a competitor page and not a routing head-to-head; those live on the comparison registry this post links to.
1. BLUF: the layer above raw speech
Raw STT and raw TTS became commodities priced per second and per character. The 2026 decision lives one layer up, in how the pieces are assembled and who operates the result. Four shapes exist:
- The component vendors. Deepgram sells recognition and the Flux agent endpoint; Cartesia sells Sonic synthesis and an interactive voice layer. You buy a capability, wire it yourself, and own everything around it.
- The open-source orchestrators. Pipecat, maintained by Daily, is a framework of pipes and services: you plug an STT service, a language model, and a TTS service together and ship the pipeline yourself. LiveKit Agents is the same idea on LiveKit's realtime transport, and the Orbit vs LiveKit and Orbit vs Agora guides cover that transport side in depth.
- The end-to-end specialists. Vapi, Retell, Bland, and Synthflow assemble STT, model, and TTS behind one console and one per-minute rate. Our Orbit vs Vapi, Orbit vs Retell, Orbit vs Bland, and Orbit vs Synthflow pages score each of them head-to-head, and Vapi vs ElevenLabs vs Retell runs the specialist round-up.
- The telco-owning platform. The shape that also owns the telephone network side: numbers, SIP trunks, carrier routing, and compliance, with the voice agent as one surface on a wider communications platform. That is Devotel Orbit's lane, and section 5 states exactly where it picks up.
The "deepgram vs cartesia vs pipecat" query that reaches this page is really a category confusion worth resolving first: those three are not alternatives to each other. Deepgram is an ear, Cartesia is a mouth, Pipecat is a nervous system. A self-assembled stack uses all three at once. The real alternative to the self-assembled stack is the specialist, and the real alternative to both is the platform that also carries the telephony.
2. Three lanes on one spectrum
Line the four shapes up and they collapse into three lanes a buyer can actually sign for.
Self-assembled. LiveKit or Pipecat for orchestration, Deepgram for STT, Cartesia for TTS, your own SIP trunk for telephony. Maximum custody, maximum engineering payroll. The stack is three vendor contracts plus a carrier, and every seam between them is yours to monitor.
End-to-end specialist. Vapi, Retell, Bland, or Synthflow. One console, one per-minute rate, telephony bundled or brokered. Custody trades away: the vendor's STT and TTS catalog is your catalog, and the latency budget is theirs to keep.
Shipped platform. The agent arrives inside a communications platform that already owns numbers, routing, and compliance, so the voice layer inherits telephony instead of bolting onto it. Custody of the model layer is expressed as tenant configuration rather than vendor contracts.
The spectrum is not good-vs-bad. It is an engineering-time-vs-custody trade: every step toward the self-assembled end buys you control and spends you a senior RTC engineer. The failure mode we see in evaluations is picking a lane for the demo and discovering at go-live that the telco layer was nobody's job.
3. The custody table: eight rows to sign before pricing
Price is the last row of this table, not the first. The first eight rows decide what you are buying.
| Row | Self-assembled (LiveKit or Pipecat + Deepgram + Cartesia) | End-to-end specialist (Vapi, Retell, Bland, Synthflow) | Devotel Orbit |
|---|---|---|---|
| TTS custody | Your contract, your vendor, your fallback chain | Vendor catalog, vendor's swaps | Tenant-owned presets, including custom voice clones |
| STT custody | Your key and your model choice | Vendor default, BYO on some plans | Tenant-sampled choice in the STT playground, persisted per agent |
| Latency budget ownership | You measure every stage yourself | Vendor-owned, methodology rarely published | Published per-stage budget and methodology, per-agent thresholds you set |
| Telephony treatment | Your SIP trunk, your bridges, your failover | Bundled or brokered, opaque to you | Native: numbers, trunking, and routing are the platform's home ground |
| Live telco ops (ports, number inventory, abuse and fraud controls) | Entirely yours | Thin or absent | Tenant-owned in the dashboard, same surface as your human traffic |
| Pricing unit | Per second, per character, per minute, three bills | Per minute, sometimes credits | Pay-as-you-go usage on one bill across voice and every other channel |
| Prompt-asset portability | Lives in your repo, fully portable | Lives in the vendor console | Tenant asset, versioned with the agent |
| Post-call data egress | You build the pipeline | Webhooks and exports vary by vendor | Transcripts, summaries, and sentiment in platform and exportable |
Read the table vertically and the lanes separate cleanly. The self-assembled column is a job description: you own every row, which is worth exactly as much as the team you staff on it. The specialist column concentrates rows two through four behind one rate, and the buyer's question becomes what the vendor does with the rows it owns. The platform column is the only one where rows four and five, the telco rows, are not an integration project.
4. Per archetype: where each lane leads
The SDK-first team. You have RTC engineers and you want the transport in your own process. Start with the transport decision: Orbit vs LiveKit for the LiveKit shape and Orbit vs Agora for the Agora shape, then add LiveKit Agents or Pipecat on top, Deepgram for the ear, Cartesia for the mouth. Budget for the rows the SDK does not cover: numbers, carrier failover, compliance, and the QA loop you will build yourself.
The open-source assembler. Pipecat is the honest choice when your pipeline is unusual: custom turn-taking, a private model endpoint, a non-standard TTS blend. Daily maintains it actively and the service interface is clean. The assembler's bill is not the vendors' invoices; it is staging, observability, telephony, and the on-call rotation, so price those before the first sprint.
The specialist buyer. You want the agent as a product and the stack as somebody else's problem. Run the four head-to-heads, Orbit vs Vapi, Orbit vs Retell, Orbit vs Bland, Orbit vs Synthflow, plus the specialist framework round-up, and score the evaluation checklist they share: latency budget, QA scoring, and training coverage. The specialists are at parity on the voice core; they separate on what surrounds it.
Every lane is defensible. What is not defensible is choosing a lane without naming who owns rows four and five of the custody table, because that is the part that pages someone at 2 a.m.
5. Where Devotel Orbit picks up
Orbit's claim is the end-to-end lane plus the telco rows the other lanes outsource or ignore: numbers, trunking, carrier routing, compliance, and the channels beside voice, on one usage bill. Within that, the model layer stays tenant-owned, and two shipped surfaces are where the self-assembled lane's payoff (model choice, sampled rather than assumed) survives inside a platform:
- The STT playground lets a tenant run its own prompts and knowledge-base audio against recognition models side by side and pick per workload; the playbook covers the method. That is STT custody expressed as a dashboard workflow instead of a vendor contract.
- The model presets page in the agents dashboard offers one-click curated presets, Balanced through Ultra Fast and Cost Saver, each annotated with its estimated latency, cost, and quality, so the latency-vs-cost dial is a tenant choice with readable tradeoffs, not a vendor's default.
- Custom voice clones keep the TTS row tenant-owned too: the voice is an asset you carry, per the clones explainer.
In other words, the rows where self-assembly pays (custody of ear, mouth, and budget) are exactly the rows Orbit exposes as tenant configuration, and the rows where self-assembly hurts (telephony, compliance, operations) are the rows the platform already runs.
6. What this post does not claim
No embedded Pipecat or Deepgram integration ships in Orbit today, and this post is not a roadmap promise of one. It documents the buyer landscape as it stands in 2026 and points, for each lane, at the shipped surface a buyer can evaluate now: the comparison registry for the specialist lane, the playground and presets for the model-custody lane, and the telephony surface for the rows nobody else owns. If your team's differentiator is the orchestration code itself, self-assemble; the stack above is real and the vendors are serious. If your differentiator is what the call accomplishes, buy the lane that ships the telco and the tenant controls with it.
Related reading
- Orbit vs Vapi
- Orbit vs Retell
- Orbit vs Bland
- Orbit vs Synthflow
- Orbit vs LiveKit and Orbit vs Agora
- Vapi vs ElevenLabs vs Retell: the framework round-up
- STT playground and knowledge bases
- Voice clones and custom TTS
Frequently asked questions
Are Deepgram, Cartesia, and Pipecat alternatives to each other?
No. They sit at different layers: recognition, synthesis, and orchestration. A self-assembled stack typically subscribes to the first two and runs the third, so the "vs" in the query is really a stack-decomposition question, answered by the custody table in section 3.
Who should self-assemble with LiveKit, Deepgram, and Cartesia?
A team whose product is the pipeline: unusual turn-taking, private models, or a transport requirement no specialist supports. The entry price is an RTC engineer plus ownership of telephony, QA, and compliance. Everyone else is buying the stack, not building it, and should evaluate the specialist or platform lanes instead.
Who owns the latency budget in each lane?
In the self-assembled lane you do, stage by stage, with your own instrumentation. In the specialist lane the vendor does, and the methodology is rarely published, which is why the round-up treats a published per-stage budget as a scoring row. In the platform lane the budget is published and the per-agent thresholds are a tenant setting you can tighten.
Can my prompts and voices move between lanes?
Prompts move as easily as they are portable: fully in the self-assembled lane, partially out of a specialist console, and as versioned tenant assets in the platform lane. Voices are stickier; a cloned voice is a vendor-specific asset, which is precisely why the clones explainer treats voice custody as its own row.
Where does telephony fit in this decomposition?
It is the row the decomposition usually forgets. STT, TTS, and orchestration vendors do not carry your numbers, ports, or fraud controls, and neither do most specialists. Before signing any lane, name who operates those rows; the custody table's fourth and fifth rows exist because that is where go-lives stall.
The takeaway
The 2026 AI-voice infrastructure decision is a custody decision: which layers you own, which you rent, and who holds the latency budget and the telephone network while you do. Self-assembly with LiveKit or Pipecat, Deepgram, and Cartesia is a real and serious lane for teams whose product is the pipeline; the Vapi family is the right buy when the agent is the product; and the platform lane wins when the call is one surface on a wider communications operation, because the telco rows come with it instead of as a project. Score the eight rows before you score the price, and the lane selects itself.
Published 12 October 2026.