We built our own runtime, and we say plainly which parts we buy.
We built the call orchestration and voice-activity detection ourselves. No third-party voice-agent platform sits in between. Where a product uses an outside provider for speech recognition or synthesis, we name that provider here, product by product, once it is verified.
Runtime specification · as of 2026-10-13
- Call orchestration
- Built in-house
- Voice-activity detection (VAD)
- Built in-house
- Third-party voice-agent platform
- None
- Speech recognition · synthesis providers
- Named per product once verified · none listed yet
- Pipelines
- One per language
- Call languages
- Korean · English · Japanese · Chinese · Bengali (varies by product)
- Deployment
- Cloud · on-premises · hybrid
- Handed to a person
- Negotiation · complaints · out-of-scope requests
On-premises installs put everything inside this line on your servers
- POS
- PMS
- CRM
- Calendar
- SMS · KakaoTalk
Negotiation · complaints · out-of-scope requests → a person
Each turn is drawn in two runs: what came in above, what went out below.
Every time the agent answers, one bar is drawn. The upper run is speech-to-text, then the model deciding the reply. The lower run is the tool calls, such as a menu lookup or an order entry, then speech-out.
How to read it
- 1Upper run · what came in
- Speech-to-text turns the caller's words into text, then the model decides the reply and the next action.
- 2Lower run · what went out
- Tool calls do the work, such as a menu lookup, a price or an order entry, then speech-out says the reply. A turn with no tool call (T3) has speech-out only.
- 3Hub 2 × 4 · turn in flight
- Where the upper run ends and the lower run begins. It turns blue only while that turn is being processed on a live call, then returns to the color of the other cells.
- 4Length · scale
- A synthetic script has no measurements, so its lengths are fixed. Once real call logs are connected, the drawing is made at 1 cell = 10 ms, with milliseconds printed on every turn.0100 ms1 cell = 10 ms · once real logs are connected
The call demo on the home page draws the same trace, turn by turn.
Open the call demoLatency gets a number only after we measure it.
We publish p50 and p95 in milliseconds for each language. Only numbers measured on our internal test bench are used, each with its date and test conditions. Until then we make no speed claim, not even “answers in under a second.”
Publishing rules
- Metric
- p50 and p95, in milliseconds
- Split by
- Language and provider
- Measured on
- Our internal test bench
- Published with
- Date · calls measured · line type · server location
- Broken down by
- Speech-to-text · model · tools · speech-out, as in the trace
| Language | Used by | p50 · ms | p95 · ms | Measured |
|---|---|---|---|---|
| KOKorean | Orda · gohotel.io · Overture AI · Sori AI · Dallang · Property OS | No measurement | No measurement | Not yet measured |
| ENEnglish | Orda · gohotel.io · Overture AI · Sori AI · Property OS · Helm | No measurement | No measurement | Not yet measured |
| JAJapanese | gohotel.io | No measurement | No measurement | Not yet measured |
| ZHChinese | gohotel.io | No measurement | No measurement | Not yet measured |
| BNBengali | Sori AI | No measurement | No measurement | Not yet measured |
None of the five languages has been measured yet. Each one gets its numbers and date here as soon as it is.
The runtime deploys three ways. Here is what runs where and what leaves your premises.
The same runtime runs in the cloud, on your own servers, or split between the two. If your security rules say call audio cannot leave the building, it installs on-premises.
| Model | Runs on your servers | Runs outside | What leaves your premises |
|---|---|---|---|
| Cloud | Nothing. Only your phone line and the systems we connect to stay with you. | Speech-to-text, the model and its tool calls, speech-out, call logs | Call audio, transcripts, tool-call contents |
| On-premises | The whole runtime: speech-to-text, the model and its tool calls, speech-out, call logs | Only tools you connect to outside services, such as SMS | Only what those outside tools are sent |
| Hybrid | The stages you choose to keep inside | The rest | Whatever the outside stages process |
Which models a product supports varies, and each product sheet says so. Dallang, for example, ships as SaaS or on-premises. For hybrid, we decide together in the working session which stages stay inside, against your security rules.
Tool calls made during a call write to these systems.
Orders go to the POS, bookings to the calendar, confirmations out by text or messenger. We confirm the product and version of each system you run in the working session.
| System | What gets written | Products |
|---|---|---|
| POS | The priced order | Orda |
| PMS | Guest requests, sorted by department (your PMS connection is reviewed in the working session) | gohotel.io |
| CRM | The caller and what they asked for (Overture AI writes to its built-in CRM) | Overture AISori AI |
| Calendar | The booked slot | Sori AI |
| SMS | The booking confirmation | Sori AI |
| The booking confirmation | Sori AI | |
| KakaoTalk | The booking confirmation | Sori AI |
Book a working session to fit it to your operation.
Tell us the deployment model, the call languages and the systems to connect, and we send an agenda built around them.