PROTOCOL NOTES / 07

Voice & forwarding

Preserve the voice packet. Keep local playback independent.

Android and iOS software implemented · Apple Wi-Fi Aware, Bluetooth LE and mixing tested on hosts and emulators only · Physical qualification pending

Authenticate once, then separate the work

Authenticated ingress splits into independent playback and forwarding work.
Authenticated ingress splits into independent playback and forwarding work. · Open full size ↗

Before forwarding, an incoming datagram must pass active-role authorization, hop authentication/replay, source SRTP authentication/replay, and the hop-envelope age bound. Source-clock freshness governs local playback only, so one relay’s view of a source clock cannot silence its whole subtree. Source decryption produces a separate clear Opus copy for local playback; the original protected bytes continue to other branches without a decode/re-encode cycle. Local encoder or renderer delays must not stall authenticated forwarding or control.

Two protection layers serve different purposes

Control and media encapsulation: source protection is preserved across forwarding hops.
Control and media encapsulation: source protection is preserved across forwarding hops. · Open full size ↗

A source uses SRTP AES-GCM around RTP/Opus. Each hop adds a directional authenticated envelope (version 3): the ride tag and hop sequence once per datagram, then for each of 1–8 bundled frames its own hop budget, age and length, all under one truncated HMAC. Relays preserve source ciphertext while updating hop-specific fields and authentication.

A link carrying several sources may hold its first frame for up to 5 ms so that concurrent frames share one datagram; a single-source link never waits. Media UDP is marked DSCP EF; whether a Wi-Fi driver honours that marking is device-specific.

Wi-Fi Aware links forward every source unchanged, including mixes received over Bluetooth LE. Over a forwarding Bluetooth LE link each side sends only its own bridge mix.

Bounded delay without cutting words

BoundMeaning
20 ms / 48 kHzSource Opus frame duration and RTP clock
1–8 framesPer bundled link datagram
48 packets per sourceBounded ingress
32 packets per linkEgress credits
40 ms / 100 msEgress queue time-to-live on Wi-Fi Aware / Bluetooth LE links
180 msForwarding age ceiling on the wire
40–160 msAdaptive playout target
600 msLate-speech horizon for local playback

A stalled radio can deliver several hundred milliseconds of speech at once. Speech up to 600 ms late still plays, as a real-time voice engine does; pauses are shortened, and playback runs at most 4% faster during speech (12% in quiet passages) until the delay is recovered. Local playout drops speech only beyond that horizon, and every such drop is counted.

Independent bounded queues prevent unbounded speech backlog. Immutable source payloads are shared across branches, and each link’s sender bundles what is queued for it into one datagram per pass. Backpressure, queue expiry, and revoked link generations are observable events. A software enqueue/send benchmark measures that stage, not mouth-to-ear latency.

Know who holds the source keys

All admitted members receive source symmetric keys. Source SRTP therefore does not prevent a malicious admitted rider from impersonating another source. Membership and transport authentication narrow the boundary, but the model is a trusted admitted group, not mutually distrustful end users.

A bridge decrypts, mixes and re-encrypts the voices of its side. It is an admitted member that holds those keys anyway, and its mix authenticates only as coming from an admitted key holder.

Authority: Media ownership; wire reference. Public examples are synthetic. This guide is maintained against the private implementation; it does not imply access to its source.