AI applications

1000

Keypad

AI applications

AI applications with their own numbers

An AI application answers a call, does its job and passes the call on to a destination. Give it an internal number to dial it directly, or choose it as a destination in IVRs, inbound routes and time conditions ("AI application"). Announcement with values: a text read to the caller by a voice model of this server, with values in it — {caller_number}, {caller_name}, {did} come from the call, any other {name} from your own system through the lookup address. Changes take effect with Apply.

The local AI service is not running: calls go straight to the destination until it runs. Local models

Application	Number	Afterwards	Status
After-hours receptionist Thank you for calling Northwind.	7502	voicemail: 1001	Active
Callback taker Tell us your name and number.	7503	🛑 End the call	Active
Order status bot Your order {order} is {status}.	7501	🎧 Queue: Technical support (support)	Active

New application

Type

Name

Internal number Optional; dial it from an extension to test.

Voice

Speed

Text Values from the call: {caller_number}, {caller_name}, {did}. Any other {name} is taken from the answer of the lookup address. Numbers, amounts and dates are read out properly.

Lookup address (optional) The server calls this address with ?caller=…&did=…&app=… and expects a JSON object, e.g. {"amount": "1.234,50"}. At most 2 seconds; the server itself makes the request.

Text when the lookup fails (optional) Read instead when the address does not answer in time or a value is missing. Empty: the text is read without the missing values.

Greeting

When it did not understand

Before passing the call on

Speech to text Language: used by cloud providers. A cloud provider receives the caller's speech (privacy).

Listen for at most (seconds)

Ask again (times)

Understanding The local model chooses well between the requests but rarely says "none of them". A cloud language model (AI → Cloud services) also recognises callers who want something else and sends them to the destination below; if it does not answer in time, the local matching decides.

Certainty needed 0–1; below it the caller is asked again and then passed on. 0.5: the keywords decide; about 0.3 lets the example sentences help too.

RequestsDestinations

For each request: keywords (word beginnings, e.g. "temiz" also matches "temizlik") — the strongest signal —, a few example sentences, the spoken reply, an e-mail address that gets the transcript and the recording, and where the call goes afterwards (empty: it ends after the reply).For each destination: a name (e.g. Accounting), keywords and two or three sentences callers would say, and where the call goes. When the caller is not understood well enough, they are asked again, then the call goes to the destination below (e.g. the keypad menu or the operator).

Example sentences are compared only by their letters. To understand callers who use other words, download the sentence-meaning model: Local models

AfterwardsWhen not understood

Calls at the same time Above this, callers go straight to the destination.

Active

Phone Settings

AI applications with their own numbers

Open Source License & Telecom Disclaimer