SingaporePanamaHong KongChina
Aldaria shows a reference architecture for private voice: GPUs, telephony and models that a deployment team can operate. We deploy agents for companies that cannot hand the call to someone else's cloud.
Operating footprint
Our stack marks · no client logos
These are template stack marks, not client endorsements. No client names, call flows or third-party logos are displayed.
A voice agent is not a prompt. It is a chain of real-time systems. When one link is rented, the latency, the data path and the failure mode belong to someone else.
ASR and TTS are designed to be trained and served on private accelerators. Domain vocabulary, code-switching and accent handling are tuned per deployment, not borrowed from a generic model.
Streaming ASR · Neural TTS · Custom lexicons
We operate the SIP edge, media path, number inventory and carrier interconnects. A dropped call can be traced without opening a vendor ticket.
SIP · WebRTC · PSTN · Porting
Current-generation accelerators in Singapore, sized for inference. Continuous batching, quantisation and KV-cache control; single-tenant pools when residency requires it.
GPU pool · Private regions · On-prem
A conversation is a deadline. Every layer you rent is a layer you cannot see.
The reference stack: SIP edge, media, ASR, orchestration, tools, TTS and playout. The template makes every hop visible so the budget can be published.
See the architectureA voice turn is a real-time workload. Inference is designed for private GPU pools in Singapore, with private weights and capacity sized for the call mix.
GPU / inference
We do not send a live conversation into a shared training queue. Speech and agent models are served from private accelerators, with a latency budget attached to every deployment.
Capacity is part of the contract. Dedicated and sovereign deployments get their own pool, their own weights and a written runbook.
Illustrative pool. GPU models, counts and capacity are agreed per deployment.
A caller does not wait for a pipeline. They wait for the next sound. This is the path our voice agents take, and the budget each stage gets.
Termination, jitter buffer, SRTP. The call enters on a private media path.
Acoustic and semantic endpointing. Barge-in detection runs in the same frame.
Streaming recognition is already warm. This is the final pass, not the first one.
Policy, dialogue state and tool dispatch. First token, not the full answer.
Speech begins while the rest of the sentence is still being generated.
Jitter control and concealment. The line stays continuous when a packet is late.
| Stage | Budget |
|---|---|
| SIP edge | 20 ms |
| Endpointing | 140 ms |
| Final ASR | 60 ms |
| Agent first token | 110 ms |
| TTS first audio | 100 ms |
| Playout | 50 ms |
| Turn | 480 ms |
Tool calls run in parallel with speech. A blocking lookup pushes the turn to about 730 ms; we design around it rather than hiding it.
Engineering targets, not marketing benchmarks. Every deployment is measured against its own budget.
Same software. Same agent. The control plane, data path and model weights move to wherever the call is allowed to live.
Shared control plane, isolated runtime. We run the models, the telephony and the pager. You run the conversation design with us.
Typical lead time: 2–4 weeks
A single-tenant GPU pool with its own capacity, routing and retention rules. Your region, your numbers, or ours.
Typical lead time: 4–8 weeks
On-prem or air-gapped. Model weights shipped, no egress, your hardware and your telephony. We stay accountable for the stack.
Typical lead time: 8–16 weeks
Every deployment starts with a written architecture decision record and a runbook. Scoping is done with an engineer, in writing. Request an architecture review
The work is rarely glamorous. It is authentication, rescheduling, dispatch, and the thirty per cent a human should take.
Authentication, order status, refunds, returns and escalation. The agent handles the repeatable call; your team gets the context and the caller.
02Scheduling, reminders, rescheduling and no-show recovery. Consent is captured at the start of every outbound call, not buried in a form.
03Triage, address capture, technician ETA and follow-up. Tool calls into the systems you already run, with timeouts and fallbacks.
04Financial services, healthcare and utilities. Retention rules, disclosure and human escalation configured per jurisdiction.
What we do not build: robocalls, political persuasion, emergency services, or a cloned voice without documented consent. We do not display client names, call flows or volumes.
Aldaria is an independent voice infrastructure company. We build the call path, the agent and the operating layer around them, so the whole system has one accountable owner.
Jurisdiction names are public. Customer contracts are with Aldaria. Public contact details, when supplied, are shown on the contact page.
One operating group. Jurisdiction names are published here; the rest of the legal record stays private.
The companyOccasional writing on latency, telephony and the design of systems that talk. Published when there is something worth writing down.
A resold voice API hides the line. The line is where calls fail, and where the latency budget is won.
A turn is a budget, not a benchmark. Where the time goes, what we refuse to rent, and what we measure.
Disclosure is not a checkbox. It is a designed moment in the call, and it changes how the agent opens.
Start a conversation
Tell us what the call has to do. We will tell you what it takes to run it.
