An AI voice app that stalls before playback may have a network establishment problem rather than slow inference. IPv6 and IPv4 fallback needs careful measurement because connection racing and real-time media setup are not the same process. This runbook separates those failure domains and defines reproducible acceptance checks.
Trace the network connection timeline
A mobile AI voice session commonly crosses changing Wi-Fi and cellular networks. The first connection may resolve an API hostname to both IPv6 and IPv4 addresses, but those routes can have different actual reachability. RFC 8305, Happy Eyeballs Version 2, published in December 2017, describes a strategy for avoiding long delays when the preferred connection family is not working. The algorithm is about connection establishment, not the quality of video once a WebRTC media flow is active. Begin with a timeline from DNS answer to socket connection, TLS setup, signaling, ICE negotiation and the first intelligible audio packet. Without separate timestamps, a slow connection may be wrongly blamed on a model inference provider. A successful browser homepage also does not establish that the API hostname or TURN relay works.
Use supported client connection racing
A dual-stack HTTP client should normally use the operating system and networking library’s supported connection racing instead of building a second ad hoc race in application code. Examine how each client handles A and AAAA records, the ordering of candidate addresses, the delay before trying alternatives, and the condition that cancels a losing attempt. The exact choices can change with runtime, platform and OS version. A Flutter application that uses one HTTP library for chat and another native networking component for video may have different selection behavior. Log the address family and elapsed time in controlled test sessions, but avoid collecting complete individual IP addresses as ordinary product analytics. Record whether a successful fallback created duplicate authentication or session-init requests; connection recovery should not become a second billable action.
Separate HTTP signaling from WebRTC media
For live AI avatars, signaling and media are separate layers. An HTTPS connection can establish quickly even when audio fails because ICE cannot reach a candidate path or TURN relay. RFC 8445 describes interactive connectivity establishment for choosing media paths; it is not replaced by the HTTP client’s Happy Eyeballs behavior. Inspect the selected candidate pair, available packet loss and round-trip time data, and whether the flow is direct or relayed. A user who sees a connected screen with no audio could be facing microphone permission, server inference, encoded audio routing or media transport. Do not conclude that IPv6 was responsible just because the problem occurred on cellular. Offer text fallback with clear status while preserving a single conversation event identifier.
Test an explicit dual-stack failure matrix
Build a fault matrix that covers IPv4-only DNS, IPv6-only DNS, dual-stack DNS, IPv6 blackhole, IPv4 blackhole and slow resolver responses. Extend it to captive portals, network switching, app backgrounding and stale DNS caches. With controlled test hosts, block one transport family and observe the actual client’s fallback latency and success. For every trial, record connection start, first successful socket, TLS completion, signaling completion and first playable remote audio. Compare p50, p95 and outright failures, not just one best-case speed. Test a session already in progress as the network changes: the system must not create a second session, re-charge the same action or silently stream two model responses. A stable acceptance test uses consistent devices and test content.
Measure regional experience without guessing
A regional comparison should be grounded in real field measurements from comparable devices, network conditions and endpoints. A synthetic 20 ms improvement in a handshake cannot be described as a 20 ms reduction in full AI conversation latency when model queues and audio synthesis dominate the end-to-end path. Report geography only at an aggregated level that respects user privacy. Compare endpoint families and route selections without inventing a ranking of countries or providers. Factor in traffic through CDN, load balancer and TURN infrastructure: the DNS region serving an API does not necessarily indicate where audio is processed. A measured fall in connection failures matters more than a tiny median improvement if it affects users trying to initiate their first interaction.
Gate releases on correctness and recovery
Release any network behavior changes behind a regional feature flag, with a documented rollback to the last proven configuration. The gate should show that dual-stack connection failures do not rise, retry attempts remain bounded, and no additional billable chat or video sessions occur after fallback. Re-run the same matrix after changing the DNS provider, TLS termination, client HTTP library or TURN service. Give incident responders a runbook that distinguishes DNS, socket, TLS, signaling, ICE and inference failures. RFC 8305 and RFC 8445 describe different layers; engineering dashboards should preserve that distinction. The outcome is a call that connects once, survives ordinary handoffs and fails with a comprehensible recovery path rather than an unexplained spinner.
Happy Eyeballs is useful when its effect is measured at the layer it actually controls. Keep session state and billing idempotent across a fallback, and publish no unsupported regional latency claims.