Media anchoring and recording
Every call is anchored and recorded from the first ring, so nothing in the conversation is lost before the agent answers.
Deze pagina is voorlopig alleen in het Engels beschikbaar. Naar de Engelse pagina
Infrastructure
Voice AI infrastructure for reliable production calls
A voice agent is only as good as the infrastructure under it. Dring brings telephony, media edge, model routing and observability together, with active and standby paths for continuity and deployment-specific controls for data and retention.
Regions
Active and standby paths, automatic failover
Calls are anchored in an active edge path, with a standby path carrying the same media anchoring, recording, caller identity and routing configuration. Exact resilience behaviour is confirmed for the deployment and carrier setup being scoped.
If the active path degrades, traffic can move to the standby path automatically. Calls already in progress on the failed leg may be affected; new calls continue on the healthy path where the configured failover supports it.
Primary pathActive
Media anchoring, recording, caller identityLive
Standby pathReady
FailoverAutomatic
The media edge
Every call is handled at the edge, before the agent speaks
The edge is where a call is anchored, recorded, identified and routed to the right tenant, ahead of anything the agent brain does.
Caller identity
Inbound callers are matched to your CRM by number before the agent speaks. Unknown numbers get the verification flow you define.
Barge-in and voice activity detection
Callers can interrupt the agent mid-sentence. Silence and voicemail are detected on outbound calls and handled by policy.
Calling windows and re-dial rules
Permitted hours per country and time zone, retry count, spacing and daily caps are enforced before a number is dialed, not after.
Tenant routing
Every call and message is routed to the correct tenant's agent, policies and data at the edge, before it reaches the pipeline.
Masking at source
Personal data is masked in the copy that gets stored: recordings, transcripts, reports and exports never carry it in the clear. See Voice Anonymizer.
Model layer
Swappable models, a fixed latency budget
The speech-to-text, language and text-to-speech models behind an agent are chosen for the language and use case, and can be swapped without a rebuild.
Swappable per agent
Speech-to-text, language model and text-to-speech providers are each selected per agent from leading vendors, and changed without touching the rest of the stack.
Turn-taking model
A dedicated model paces the conversation: real pauses, interruption handling and natural turn-taking instead of a flat, scripted read.
Deployment-specific processing
Processing location, transfer controls and data boundaries are documented and reviewed against the requirements of the deployment.
10K+agent conversation minutes/day
Observability and operations
Watched around the clock, by people who answer
Infrastructure is a service, not a shipment. Monitoring, alerts, staged rollouts and a named incident playbook come with every line.
Health checks
Every trunk, SIP edge and media pod is probed continuously, so a line that is quietly failing is found by a check rather than by a customer.
Alerts
Degradation pages the on-call engineer before customers notice, whether it is a trunk losing calls, a SIP edge refusing registrations or a media pod struggling.
Auto-scale
Media capacity follows concurrent calls, up and down, so a campaign that doubles its dialing needs no hardware order and quiet hours do not pay for peak.
- Monitor
Live line health
Answer rates, audio quality signals and latency, watched continuously. Alerts go to Dring first, then to you if action is needed.
- Roll out
Controlled stages
Changes to an agent or route go live in controlled stages, with the previous version available for rollback.
- Respond
Incident playbook
Defined severity levels, a named owner, customer communication within the first hour of a confirmed incident.
- Regress
Weekly regression sweep
Every live agent is swept weekly against the previous release, catching drift before a customer notices it.
Data and isolation
Isolated by tenant, encrypted end to end
Data belonging to one tenant is never reachable from another, and every hop, in transit and at rest, is encrypted.
Tenant isolation
Enforced across the pipeline. Transcripts and recordings improve only your own agent, never anyone else's.
Encryption in transit and at rest
Data is encrypted in transit with TLS 1.2 or higher, and encrypted at rest throughout the pipeline.
Retention per contract
Retention defaults apply unless your contract sets its own, with deletion available on request per number or per customer.
Full detail on data flow, retention periods and data subject rights is on the security page. See telephony for the call path, the Agent Factory for the operating loop, the voice AI operations snapshot for context and the 62-language customer service guide for language planning.
Technical flow
A call's path through the infrastructure
From the edge to the model layer, post-call analysis and your systems, every stage is instrumented.
FAQ
Infrastructure questions
What happens if an edge path has an outage?+
Traffic can move to the configured standby path automatically. Calls in progress on the failed leg may be affected; new calls continue on the healthy path where the deployment supports it.
How is processing location agreed?+
Processing location, transfers, retention and access are reviewed against the workflow, contract and applicable data requirements before production.
How are changes to an agent or a route rolled out?+
In controlled stages, with scores watched at each step and the previous version available for rollback.
Is our data isolated from other tenants?+
Yes. Tenant isolation is enforced across the pipeline, and transcripts are used only to improve your own agent. See security for the full detail.
Plan your production voice AI setup
Talk through telephony, failover, retention and where your data is processed. A consented form submission can be followed by an AI qualification callback in about two minutes, with timing confirmed for the workflow.