Voice AI: what is built and what has not happened yet

Can the platform speak on the telephone, and what state is it in

The state in one line

The voice channel is built: seven tables with company isolation, its own audio stream gateway, telephony, speech recognition and speech synthesis adapters, and a screen in the panel. Not a single real call has been made, and until the first one the channel is not counted as working.

That is not careful wording. Telephony is connected by contract: a number, provider accounts, keys and a public deployment address are required. Until every one of those is issued, the panel lists what is missing and refuses to put the channel on a real line.

What the channel does

Everything listed here is written and covered by tests without a line. Not one line below is a plan.

• Inbound call: accepting the telephony notification, answering with markup and attaching a two-way audio stream.
• Outbound call: started by a person from the panel, with a pre-flight check. There is no dialling from a list and none is planned in this phase.
• Conversation: streaming speech recognition and streaming synthesis, barge-in that stops its own speech, and protection against hearing its own voice.
• Answers from the knowledge base: the same grounding check as in written chat. A claim without a record behind it is not spoken — «no data» is spoken instead.
• A transcript of the conversation, a call journal, and handover to a person carrying the promises made and the questions left unanswered.
• A do-not-call list, consent to recording, and disclosure that a machine is speaking.
• Spend ceilings: the cost of the call and the daily ceiling of the organisation are checked BEFORE the reply, not on the invoice.
• A demonstration conversation in the panel that walks the same path as a live line and is labelled as a demonstration.

What we do not promise

The list is short and exact, because every line in it is something that cannot be checked without a real call.

• That the assistant has already spoken to a live caller. It has not.
• Any latency figure «on the phone»: only internal steps were measured, not a conversation.
• The quality of Kazakh in voice: the path works but has not been measured on live speech.
• The share of false and missed barge-ins: those are measured on live speech only.
• The cost of a minute of conversation: provider tariffs are not confirmed by an invoice.
• Outbound calling from a list: the product has no such capability.

What we need from you for the first call to happen

Eight items, and most of them are not engineering: the owner closes them together with whoever runs the deployment. The internal list has ten — neighbouring technical items are merged here so the list reads rather than counts.

• A telephony account and the number calls go from and to.
• A public deployment address with a valid certificate.
• Keys for the recognition and synthesis providers, or consent to run on the fallback path.
• Provider tariffs: without them a live call stops under the rule «nothing to count with».
• Provider hosts on the outbound traffic allow-list.
• Three ceilings in the organisation policy: cost per call, cost per day, calls per day.
• A legal review of call recording and of the disclosure text in your jurisdiction.
• A deliberate decision to switch the line on.

What about the price of voice

The price list has a line with voice, and five limitable quantities are named for it: included conversation minutes per month, phone numbers, conversation scenarios, branches and international telephony. The owner has put a number against none of the five, and the pricing page says «not decided» in words rather than a plausible figure.

Basis: docs/VOICE_AI_IMPLEMENTATION_STATUS.md, docs/LIMITATIONS.md. This page describes what exists in this build, not what is planned.