Skip to content
Phone Guides

Asterisk AI provider contract checklist

This page records the provider contracts referenced by the Asterisk and FreePBX AI guides. It was checked against provider documentation on 20 October 2026. These are provider statements, not proof that an Asterisk or FreePBX integration works. Pin the selected API, model and SDK in the application, then recheck the linked contract before a deployment.

Source-reviewed, 20 October 2026. Phone Guides has not made provider calls or tested these provider APIs with Asterisk or FreePBX. Prices, terms and service features can change; verify the exact account and model before use.

OpenAI GPT-Live

  • API and model: Live Sessions API, gpt-live-1, generally available. The documented server WebSocket is wss://api.openai.com/v1/live/sessions; send session.start, wait for session.started, and close with session.close. Record session.closed usage when received. Startup choices include model, instructions, audio format, voice and delegation mode. See GPT-Live WebSocket guide and model page.
  • SDK: The guides describe the wire API and do not pin an OpenAI SDK. For an adapter, record the actual client library/version or the raw WebSocket implementation in its lockfile and deployment record. Do not infer SDK compatibility from the API name.
  • Audio: The Live API supports negotiated media for WebRTC and SIP. The WebSocket reference documents PCM16 and G.711 μ-law/A-law at 8 kHz when selected. Its audio goes inside JSON events as base64 bytes. Asterisk ulaw therefore needs envelope conversion even when codec conversion can be avoided. Confirm the actual session format and Asterisk MEDIA_START format.
  • Lifecycle: WebSocket audio appends have no per-append acknowledgement. The primary stream lacks audio timing fields and a per-response audio-done event; track queued and played bytes yourself. session.closed carries final usage. Active session time includes silence and backend work, and the session price is billed per second. Backend model and tool costs are separate. See session management and voice cost accounting.
  • Published price: US$0.05 per session minute, billed per second without rounding up to a whole minute, checked 20 October 2026. Backend calls and tools are charged separately. Recheck the model price at selection time.
  • Data terms: Review the API data controls for the exact endpoint, project and account eligibility. This guide does not claim a particular retention mode, regional setting or Zero Data Retention approval for a Phone Guides example.

Deepgram Voice Agent

  • API and model: Voice Agent WebSocket wss://agent.deepgram.com/v1/agent/converse. The client waits for Welcome, sends one Settings message and waits for SettingsApplied before audio. The chosen listen, think and speak providers and model strings belong in that settings record. See message flow, settings and API reference.
  • SDK: The Asterisk examples use the wire contract and pin no Deepgram SDK. Deepgram's 12 October 2026 SDK release note, checked 20 October, lists Python 7.13.0, JavaScript 5.15.0, Java 0.11.1 and Go 3.9.0. The first three add experimental custom Think-provider messages, which require a configured custom Think-provider WebSocket; Go 3.9.0 supports typed Voice Agent function requests, responses and cancellation events. These are optional client-library capabilities, not an Asterisk connector or proof the Phone Guides examples run on those versions. Pin the package actually used by an adapter.
  • Audio: Settings define encoding and sample rate. Official examples include linear16 input at 16 kHz and output at 24 kHz, and μ-law at 8 kHz both ways for a Plivo bridge. These are examples, not a guarantee for every selected model/configuration. The WebSocket carries audio as binary frames. Match the actual Asterisk media channel and provider settings; resample or transcode when they differ.
  • Lifecycle: Wait for Welcome then SettingsApplied; handle UserStartedSpeaking, function-call request/cancel events, AgentAudioDone, errors and warnings. Deepgram's 8 September 2026 changelog documents speculative function calls: they dispatch before speech recognition confirms the turn by default. Set defer_until_eot: true on each function that must wait for turn confirmation; the default is false. On FunctionCallCancelled, stop client-side work for the named function ID and send no FunctionCallResponse. The cancellation signal does not reverse a side effect already submitted to another system. Settings initialize the connection; reopen and reinitialize if a setting cannot be updated through a documented update event.
  • Published price: The pricing page, checked 20 October 2026, lists per-minute rates by plan and configuration: Standard US$0.075 pay-as-you-go / US$0.068 Growth; Standard with BYO TTS US$0.065 / US$0.051; Custom with BYO LLM and TTS US$0.050 / US$0.041; Advanced US$0.163 / US$0.146; Advanced with BYO TTS US$0.122 / US$0.110. The page displays US$0.059/min for Custom BYO LLM, but the two-plan table does not assign that single value to a plan and the Pay As You Go-only view omits it. Treat the plan-specific BYO LLM amount as Unknown until Deepgram confirms the account's price. The listed Voice Agent concurrency is up to 45 connections on pay-as-you-go and the pricing page lists up to 60 on Growth. Confirm the chosen configuration, plan, region and account quote; the prior launch-announcement price is not a current quote.
  • Data terms: The data guide says Voice Agent audio, intermediate transcripts/context and synthesized audio are retained for model improvement by default. mip_opt_out: true opts out and limits content retention to processing; usage metadata/logs remain retrievable for 90 days. Regional Voice Agent hosts include Australia, EU and India; a regional endpoint plus opt-out is needed for Deepgram's full in-region statement. A third-party LLM configured for think follows that provider's separate terms. Confirm the endpoint and setting in the actual Settings message and account.

ElevenLabs ElevenAgents SIP trunk

  • API, agent and access: ElevenLabs documents SIP trunking as a managed-agent connection for generic SIP-compatible telephony systems; its page does not name Asterisk or FreePBX. The SIP call itself has no provider-specific API version or client SDK; pin the selected agent/version and any /v1/convai management or webhook endpoint separately. The documented SIP trunk feature requires an Enterprise account or an Enterprise trial. See SIP trunking, post-call webhooks and the agent overview.
  • Audio and limits: SIP media uses PCMU or PCMA (G.711) at 8 kHz or G.722 at 16 kHz. Confirm SDP negotiation and media encryption with the exact trunk: TLS 1.2+ is documented for SIP signalling, while SRTP support varies with the connected trunk. UDP is documented as experimental for testing, not production. The provider publishes plan-specific concurrency and queueing behaviour; confirm the Enterprise account's own call, rate and duration limits rather than importing self-serve limits.
  • Lifecycle: The provider's SIP service receives or places the call and its signalling ends the dialog with BYE; agent handling and Asterisk/PBX routing remain separate state owners. The transfer_to_number SIP REFER path requires the trunk to allow REFER, and SIP REFER transfers do not support a warm message to the receiving person. If a SIP-over-TLS connection has closed, follow-up BYE or REFER routing must follow the SIP dialog's Contact/remote-domain guidance. Optional post-call webhooks can report transcription, audio or call-initiation failure; verify HMAC signatures, deduplicate retries by conversation ID and do not treat post-call delivery as live call state. No Asterisk/FreePBX connector or exact PBX transfer sequence is documented by ElevenLabs.
  • Published price: The public self-serve pricing page lists US$0.08 per additional call minute and US$0.16 per minute for enabled burst calls above concurrency; monthly plans include different call-minute and concurrency allowances. LLM usage is billed separately and the connected SIP trunk/carrier bills its own charges. The SIP-trunk feature is enterprise-gated and Enterprise pricing is custom, so those self-serve rates are not an Enterprise SIP quote. Confirm rates, metering and concurrency with the account. See ElevenAgents pricing, checked 20 October 2026.
  • Data terms: The privacy guide and retention settings, checked 20 October 2026, say conversation transcripts and recordings are retained for two years by default. Retention can be set separately for transcripts and audio, including 0 days for scheduled deletion; audio saving can also be disabled. Check the selected agent, account mode, any external LLM, SIP carrier, recording/webhook settings and their separate terms before routing calls.

Microsoft Azure AI Voice Live

  • API and model: Voice Live API 2026-07-15 is generally available and is the default service version used by Microsoft SDKs. Select and record a model and Azure resource type/region; feature support differs between Microsoft Foundry resources and Azure Speech in Foundry Tools resources. The Asterisk examples use documented raw WebSocket events and pin no Voice Live SDK. Microsoft’s 15 July release notes list Python azure-ai-voicelive 1.3.0, C# Azure.AI.VoiceLive 1.2.0, Java azure-ai-voicelive 1.1.0 and JavaScript @azure/ai-voicelive 1.1.0; these versions are provider SDK releases, not tested Asterisk adapters. See the versioned API reference, release notes and resource/model overview.
  • Audio and limits: The API reference allows g711_ulaw and g711_alaw input and output at 8 kHz; G.711 input defaults to 8 kHz. This can match an Asterisk ulaw or alaw channel without codec resampling, subject to the actual negotiated media format, but the JSON event envelope still needs translation. The API emits rate_limits.updated with configured quota, remaining quota and reset time; the numeric limits depend on the resource/account and are not a fixed general value in this checklist. Read expires_at from session.created; the server sets it and clients cannot extend it with session.update. See the audio format and event reference.
  • Lifecycle: Open the versioned server-to-server WebSocket, wait for session.created, configure and confirm with session.update/session.updated, then process provider-specific conversation, response, audio, tool and error events. parallel_tool_calls defaults to true; set it to false when that session must issue tool calls sequentially, and still enforce authorization, action locks and idempotency in the application. Handle output interruption/truncation and live rate_limits.updated state; close provider and Asterisk resources separately. The reference gives a server-set session expiry timestamp but no single duration that applies to all models/resources. No Asterisk connector or FreePBX integration is established by Microsoft's API contract.
  • Published price: Microsoft lists Pro, Standard, Lite and BYO tiers with separate text/audio/native-audio billing categories, model-based tier assignment and additional charges for custom speech/voice training and hosting. The public pricing table currently renders the per-unit Voice Live rates as $-; an exact usable amount is therefore Unknown from this public page. Get a region/resource-specific quote and estimate cached context, audio and any separately billed features before selection.
  • Data terms: Microsoft's Voice Live privacy and security guide, last updated 2 June 2026 and rechecked 20 October, says Voice Live itself does not retain customer data after processing. The API still processes prompts, generated output and tool results. Features it invokes, including Foundry agents and custom voice/avatar, may store data under their own terms; connected Microsoft or third-party services may process or store it outside the Azure compliance boundary. Optional support/debug logging is enabled for a support case and stores speech data in Microsoft-managed Azure storage in the resource region, with automatic deletion after 30 days and controlled access by authorised support engineers. Confirm the exact resource, model, processing scope, connected services, logging choice and enabled features.

Amazon Nova 2.5 Sonic on Bedrock

  • API, model and SDK: AWS announced Nova 2.5 Sonic as generally available on 5 October 2026. The model uses the Amazon Bedrock bidirectional streaming API; the model ID listed by AWS is amazon.nova-2-5-sonic. AWS's Nova 2 guide currently documents the shared InvokeModelWithBidirectionalStream event contract and its sample uses the separate aws_sdk_bedrock_runtime Python package with SigV4. The sample does not pin a package version. AWS lists four launch regions: us-east-1, us-west-2, eu-north-1 and ap-northeast-1. See the GA notice, Nova 2 stream guide and Bedrock model ID list.
  • Audio and limits: The bidirectional event schema accepts mono LPCM, 16-bit, at 8, 16 or 24 kHz for speech input and output. The starter sample uses 16 kHz input and 24 kHz output. An Asterisk G.711 path must be decoded to LPCM even at 8 kHz. The Nova 2 getting-started guide gives an eight-minute connection limit; the Nova 2.5 service card says the API surface is the same but does not state its own duration cap. The stream guide reports cumulative text and speech token usage. The Nova 2 Sonic model card lists 20 concurrent streams per account per supported Region and says that quota is not adjustable, but the checked AWS material does not establish that this quota also applies to Nova 2.5 Sonic. The Nova 2 guide does not state a general audio-frame size ceiling. Confirm Nova 2.5 access, duration and quota in the target Region.
  • Lifecycle: Send sessionStart, then promptStart; open content with contentStart, stream audioInput events, and close content with contentEnd. Close the prompt and session with promptEnd and sessionEnd. On barge-in, stop queued playback on the Asterisk side when the provider interruption arrives. AWS's Nova 2 guide documents connection renewal at its eight-minute limit: open a new stream and seed saved history after the system prompt and before new caller input. The Nova 2.5 card states the same API surface, but the exact duration and automatic SDK handling for 2.5 are not specified in the checked sources; verify the selected SDK/model path. AWS's guide says history can be supplied once in that position. The provider stream does not create or clean up Asterisk channels and bridges. See the input events, output events and error recovery guide.
  • Published price: AWS's 5 October release notice says Nova 2.5 Sonic uses the same pricing as Nova 2 Sonic. The public Nova and Bedrock pricing pages checked 20 October 2026 describe token-based audio/text charges and show the speech-to-speech pricing section, but do not expose a readable per-unit rate in the checked public page text. Record the exact Region, audio/text input and output rates from the account's Bedrock price view or quote before calculating call costs; the usable amount is Unknown from the public pages checked.
  • Data terms and authentication: The current Nova 2.5 Sonic service card, dated 5 October 2026, says Bedrock does not store or review prompts, caller speech or generated text/audio, does not share them with other customers or third-party model providers, and does not use Bedrock inputs or outputs to train its models. Conversation context is held in memory while the stream is active and becomes eligible for garbage collection after session teardown. Customer-configured S3 storage, CloudWatch and CloudTrail records remain separate settings to review. For authentication, AWS's current API compatibility guide and API keys reference explicitly exclude Bedrock API keys from InvokeModelWithBidirectionalStream; use standard AWS credential-based SigV4 authentication and least-privilege IAM. The Nova 2 Sonic model-card sample also shows an API key, which conflicts with those endpoint-level rules, so do not use that sample's key step for this operation. Check region, logging and the account's AWS terms.

Optional orchestration SDK: Strands Agents Bidi

  • SDK and model: The Python package strands-agents 1.61.1 was published 15 October 2026; the current Strands quickstart shows BedrockNovaSonicModel(model_id="amazon.nova-2-5-sonic"). The Nova voice quickstart requires Python 3.12+, despite the general package supporting Python 3.10+. This is an agent framework version, not a Bedrock API version or an Asterisk connector. Pin its voice extras and all resolved dependencies in the application lockfile. See the PyPI release, Bidi quickstart and custom I/O guide.
  • Audio and limits: The local AudioIO example uses PortAudio microphone/speaker devices. Custom InputStream and OutputStream handlers can accept WebSocket input, but the guides do not specify a telephone codec, Asterisk frame size or G.711/ARI adapter. The underlying Bedrock model format, region and eight-minute connection limit remain as described above.
  • Lifecycle: BidiAgent.run() remains open until cancelled, a timeout occurs or a tool calls agent.cancel(). Custom streams can initialise in start(agent) and release resources in stop(). This application lifecycle is distinct from Bedrock's model stream and Asterisk channel/bridge lifecycles. The docs do not establish automatic renewal of Nova's eight-minute connection; record and test that behaviour for the exact SDK release. See Bidi I/O streams and session management.
  • Price and data: The SDK is Apache-2.0 licensed; it adds no stated per-minute voice price. Bedrock model usage and hosting/telephony infrastructure are billed separately. Strands can persist agent conversation state with SnapshotSessionManager and configured storage; use the AWS provider terms above for Bedrock processing, and review the application storage, telemetry, recordings and transcript settings separately. The SDK package does not establish a provider data-retention term.

Google Gemini Live API

  • API and model: The WebSocket API uses v1beta and the Live API remains in Preview; the selected model gemini-3.8-live is separately marked Stable. Google lists gemini-3.1-flash-live-preview as a legacy preview and recommends migration. See the WebSocket API reference, model page and Live API capabilities.
  • SDK: The official JavaScript SDK release was @google/genai 2.30.0, released 14 October 2026. Pin the SDK actually used; the Asterisk wire-contract examples do not establish that this SDK version has been tested with Asterisk. See official SDK releases.
  • Audio and session limits: Native audio input is raw PCM16 little-endian at 16 kHz; generated audio is PCM at 24 kHz. Audio-only sessions are limited to 15 minutes and audio-plus-video sessions to 2 minutes unless session-management techniques extend them. An Asterisk 8 kHz μ-law path requires decode/resample and output conversion. See audio capabilities and session management.
  • Lifecycle: Stateful WebSocket setup and session messages must be handled. Google says the server may reset the connection: retain the newest SessionResumptionUpdate token and use it to resume; the token is valid for two hours after the last session terminates. Handle GoAway.timeLeft before the connection closes, generationComplete before advancing response work, and server_content.interrupted by immediately clearing queued client audio. Context-window compression can extend session duration. These are provider-session controls; they do not restore Asterisk channel or bridge state. See Google's Live API best practices and session-management guide.
  • Published price: For the standard paid tier, the pricing page lists gemini-3.8-live audio input at US$3.00 per million tokens (Google's estimate: about US$0.005 per audio minute) and audio output at US$12.00 per million tokens (about US$0.018 per audio minute), checked 20 October 2026. This is token billing, not a flat call rate; accumulated context, transcription text and tools can add charges. Free-tier content terms differ, so model the exact account/project use.
  • Data terms: Google's Gemini API terms say unpaid-service prompts and responses may be used to improve products and may be reviewed by people. Paid API access through a project with active Cloud Billing is treated as paid service; Google says it does not use prompts/responses to improve products and applies its Data Processing Addendum. Verify billing status, region, account terms and any separate tools before sending call data.

Google Cloud Speech-to-Text used by the AEAP example

  • API and SDK: Keep the AEAP sample's exact package lock and API endpoint distinct. The upstream sample currently locks @google-cloud/speech 4.10.0 within its ^4.9.0 range and calls Speech-to-Text v1 StreamingRecognize over gRPC. The current Node SDK changelog lists 8.2.0, with major v8 requiring Node 22; that is not a drop-in version for the sample's v4 dependency range. See the sample package, sample lock and official SDK changelog. The guide does not declare the sample broken or claim v1 has been discontinued.
  • Audio and request limits: v1 streaming contract in the example is separate from v2. The current v2 streaming guide documents a 25 KB limit for the initial request and each subsequent message. Do not apply that v2 limit to a v1 request without checking the v1 contract. Confirm encoding, sample rate, stream duration and project quotas for the chosen endpoint.
  • Lifecycle: The sample uses the v1 gRPC stream, while v2 has its own recognizer resource and request shape. A migration requires an explicit API/client port and test of stream start, interim/final results, errors, closure and reconnect behavior. A package version update alone does not establish API compatibility.
  • Published price: The Cloud Speech-to-Text pricing page lists v2 Standard recognition at US$0.016 per processed minute up to the first 500,000 minutes per account/month; it bills successful processing in one-second increments. The page has separate v1 rates and model tiers. Do not apply the v2 figure to the AEAP v1 request; price the exact API/model/account tier separately.
  • Data terms: Google Cloud API settings and terms govern Cloud Speech-to-Text, separately from Gemini Developer API terms. The setup guide documents v1 data logging as an opt-in feature. Check the exact project, endpoint, data-logging setting, region and Cloud terms; this source review does not establish the sample's runtime configuration.

Sangoma Scribe managed service

Sangoma's Scribe guide, checked 20 October 2026, describes a commercial module for FreePBX/PBXact 15, 16 and 17. Its update stamp says “Aug 12” without a year, so this is the guide's stated range and not confirmation of current release support. The manual describes transcription and analysis for voicemail and call recordings, uploaded audio transcription, reprocessing older recordings, and text-to-speech for system recordings. Sangoma's licence purchase guide lists a monthly SKU with 1,000 shared minutes and a top-up SKU with 1,000 shared minutes valid for three months; unused monthly minutes do not carry over, and top-up minutes are used after the monthly allowance. The FreePBX product page, checked 20 October 2026, lists $24.99/month for 1,000 processed minutes but does not state the currency; it does not list a top-up price. Confirm tax, currency and current account price before purchase. The manual says licence minutes are consumed for generated TTS audio and its dashboard reports usage for TTS, uploaded audio, voicemail and call recordings. TTS makes PBX prompt audio; none of these descriptions documents a live, bidirectional caller agent.

The public manual does not disclose a provider API/SDK version, maximum recording or upload length/file size, top-up price, processing region, exact audio/transcript retention or whether a customer can pass Deepgram's direct API settings through Scribe. Do not apply the Deepgram direct API settings or terms above to the managed Scribe path. Confirm current module support, purchase price and consumption rules, partner terms, region, subprocessors and retention with Sangoma for the exact FreePBX release and account.

Sangoma's published Data Processing Agreement, version 3.1 dated 11 October 2024, lists sub-processors and processing locations for named Sangoma services. Its Annex F does not name Scribe or FreePBX in the applicable-services column. Do not use that general list as evidence of Scribe's partner, region or retention terms; check the current contract and exact service path with Sangoma.

Recheck before a release

For every provider path in a deployment record, save the check date, exact model/API and SDK lock, input/output codec and sample rate, frame or message limits, setup and close events, interruption/cancellation behavior, rate limits, current price basis, account tier, processing region, retention/data-use setting and terms link. Recheck provider pricing, model availability and data controls before each publication or system release. This checklist records public documentation only; it is not a provider endorsement, legal determination, Asterisk connector test or FreePBX support statement.