Skip to content
< >contact
all products
applied AI · live

coralSpeak

A Turkish-speaking operator on your phone lines, in your own voice — and, if you want, on your own hardware.

Overview

coralSpeak connects to your switchboard over a standard SIP trunk and answers the way a person would: it greets, listens, answers from your knowledge base, and hands the call to a colleague when it should. Each agent takes an extension on your PBX, so your numbers and your call flow stay exactly as they are. Every layer — the voice, the transcription, the model — runs either on your own GPU or on managed cloud, and the choice is made per agent from the panel. One customer can be entirely local and the next entirely hosted, in the same installation, switched while it is running.

Specifications

Audience
applied AI
Platforms
Web · iOS · Android · Self-hosted
Status
live
Built with
SIP trunk · own GPU or cloud

Capabilities

Your own voice, from 15 seconds

A clean 15-second recording is enough: the announcement voice already on your switchboard, your best agent, or a studio take. Tone, tempo, feeling and speed are yours to set and to save as a preset bound to an agent. 40+ ready Turkish voices ship alongside it.

Turkish handled first, not translated into

Numbers and dates are read as words, so 1.250 TL is spoken and not spelled. The Turkish lowercase rule is applied properly — İ to i, I to ı — because the other way round the model reads nonsense. Brand and technical terms get a pronunciation dictionary, and pauses are placed at sentence ends rather than sprayed through the line.

It answers from what you gave it, and nothing else

Upload PDF, Word, Excel, HTML, CSV or plain text. Each customer sentence is matched to the passage that answers it, and the reply is grounded in that passage. With strict mode on, anything outside the knowledge base is declined politely — general knowledge, arithmetic and small talk included.

A turn in about a second and a half

Measured on real SIP calls: 0.1 of a second from the caller stopping to the transcript being ready, under 0.2 of a second to the model's first token, and about 1.2 seconds until the first audio is on the line. It was eight to nine seconds before the work that closed it.

Built for a telephone line, not a browser tab

The agent stops the moment the caller starts speaking. Room noise does not cancel the greeting. Audio is generated and streamed a sentence at a time, so even a long answer starts arriving immediately, and it speaks G.711 — the codec Turkish PSTN actually runs on.

Separation you can point at

Every customer is a workspace with its own voices, agents, knowledge base and SIP identity, separated at the application layer and at the SIP layer both. Usage is metered by the second against a ledger rather than a single editable number, and generation stops when the balance does.

Try it on your own data

Open an account on any of the 5 and test it against your own data. No per-seat licensing and no vendor lock — you pay for what you use. Our team is ready to talk about deployment, white labelling and pricing.