Encryption

Encryption

Voice protocol v3 removes its own envelope encryption layer. Confidentiality is provided exclusively by Meshtastic's channel encryption: the firmware AES-256-CTR-encrypts the entire Data payload (this protocol's header + body) using the channel PSK before handing the frame to the LoRa PHY. Peers without the channel PSK cannot read the bytes.

This means:

• There is no per-message AEAD at this layer (v2 had AES-256-GCM with an HKDF-derived per-message key; removed in v3 to save ~28 octets per chunk and simplify the surface).
• The 4-byte mac_tag trailing the header is unkeyed SHA-256(header[0..12])[..4]. It catches on-air bit-flips (AES-CTR is bit-flip malleable) but is not an authenticator — any peer with the channel PSK can forge a valid tag.
• A consequence: a malicious channel member can fabricate or alter any frame on this layer, including NACK frames with give_up=true. That matches Meshtastic's threat model for text messages on the same channel.

What about end-to-end confidentiality between two peers?

Out of scope for this protocol. If you need it, layer your own AEAD on top of the codec frames before handing them to build_message, or wait for a future revision — git history (the v2 line) shows what a per-message envelope looked like and can be revived if a concrete threat model warrants it.

Migration from v2

V2 wrapped each DATA/PARITY body with AES-256-GCM keyed by HKDF-SHA256(channel_psk, message_id ‖ from_node_num, "voicetastic/v2"). V3 drops the wrap entirely; the encrypted flag bit (0x20) and mac_keyed flag bit (0x08) are now reserved-zero and v3 parsers reject frames that set them. The wire-version byte bumps from 0x02 to 0x03 so v2 ↔ v3 frames cleanly fail the version check rather than producing silent corruption.

See VOICE_PROTOCOL.md §7 for the authoritative spec.