ShaAnas 1.0 / AudioBridge
Discuss a controlled voice integration with an AI service you already use. We first assess compatibility, agree the backend boundary, then test the selected setup privately.
Implementation access by agreementThe public preview contains five recorded synthetic scenario results. It does not analyse your traces, connect your AI or host a live call.
Three agreed steps
Tell us which AI, telephony service and language you use. We review audio formats, response control, interruption and context acknowledgements for that specific setup.
You can bring your existing AI. Creating a separate ShaAnas AI is not required.
Define an owner-controlled backend and the required API or provider adapter. The server operator manages credentials, authenticated call evidence, media access and shutdown.
The implementation is supplied privately by agreement. This page provides no public key form, universal provider endpoint or automatic account connection.
Agree access to the evaluation components, a small test scope and expected outcomes. Start with controlled checks, then test the authorised live path and review its technical results.
Your own traces and AI are evaluated within the private pilot, with the data and access boundaries agreed beforehand.
The exercised voice path uses Twilio and English. Bounded reference calls and separate local checks cover audio transport, interruption and context gates.
Other AI services, telephony providers, languages and deployment platforms require their own integration review and testing.
Qwen, Whisper and eSpeak were used in the local QA pipeline. They are reference test components, not a required AI product that you must adopt.
A chosen service needs a compatible server adapter. Supplying an arbitrary URL and key alone does not establish that compatibility.
This preview is not a ready hosted SaaS service or a general-purpose public API. Live infrastructure, provider configuration and evaluation scope are agreed for each pilot.
The interface below describes the existing server boundary for an integration discussion. It contains no implementation and is not a client kit. The implementation and its access terms are supplied privately by agreement.
The generic createLiveVoiceHost supports providerKind:'openai' or 'local' and an injected providerFactory(hooks). The local createShaAnasLocalBridge facade composes the local runtime; its reviewed dependencies can include createBackend and createProvider. These are server composition interfaces, not public hosted endpoints.
The factory receives onEvent, onSignal, onTranscript, a shared nowMs clock and a fresh localIdPrefix. It returns eight synchronous methods:
| Method | Boundary |
|---|---|
connect | Explicit cloud connection or zero-argument local start; {accepted,code} confirms a local request, not readiness. |
appendAudio | Copies admitted incoming audio synchronously; returns exact {outcome}. |
createResponse | Uses the host generation and a reserved local turn; returns {accepted,code} before asynchronous provider results. |
cancelResponse | Retires the matching turn and returns exact {outcome}. Actual cancellation/context retirement must still be observed. |
restoreContext | Requests the exact private context reference/revision; a later matching acknowledgement must confirm the real installation step. |
tick | Enforces bounded work and session deadlines using the shared owner clock. |
snapshot | Returns typed technical state; observed transport closure is separate from a request to close. |
close | Revokes readiness and late output, then closes owned resources. |
Audio/cancel outcomes are exactly local-enqueued, not-sent or unknown. Unknown effects are not automatically retried. Readiness requires an effective configured session acknowledgement, or actual loaded local models and verified runtime/audio settings. A socket opening or process starting is insufficient.
The authenticated telephony receiver supplies admitted raw μ-law audio at 8,000 Hz, mono, in chunks of up to 1,600 bytes. The adapter copies incoming bytes before returning, converts formats as needed for the chosen AI, and returns matching normalized response audio to the bound telephone output queue.
Response events carry local turn identities, generations and ordered chunks. The host keeps current call admission, provider cancellation, queued playback and telephone clear barriers separate. Provider completion and playback marks do not independently prove human hearing.
A context acknowledgement must match its exact reference and revision after the actual supported restore operation. The local reference evaluates a private prompt and explicitly resends installed history with future stateless requests. It does not claim persistent model memory or general semantic continuity. Old audio and actions are not automatically replayed.
Credentials belong in the agreed owner-controlled backend. Private transcripts and raw audio do not belong in public technical reports. Data retention and any private dialogue capture are agreed for the pilot; the public page accepts neither credentials nor conversation data.
Start with the setup you already have
Share the provider name, the intended call flow and a small test objective. Please do not include API keys, private audio or customer traces in the initial enquiry.