Voice Protocol

Voicetastic Voice Protocol

Voice messaging over the Meshtastic mesh.

This document is the source of truth for the wire format used by voicetastic-desktop and any compatible peer. The reference implementation lives in crates/voicetastic-core/src/voice/.

For an implementer-friendly companion (byte walkthroughs, diagrams, sender/receiver recipes), see Home.

Protocol version: 3 (wire byte 0x03)

V3 is wire-incompatible with V2: the body envelope encryption (AES-256-GCM with a per-message HKDF-derived key) and the keyed-MAC variant of the trailing header tag have been removed in favour of Meshtastic's channel encryption. The four-byte trailing header tag is now always unkeyed SHA-256(header[0..12])[..4] and serves as a tamper-detection check (channel AES-CTR is bit-flip malleable, so this remains useful even though the payload is encrypted on the wire). A V3 parser MUST reject any frame whose version byte is 0x02.

1. Goals

Carry recorded voice (any narrowband codec) over Meshtastic LoRa packets.
Survive packet loss without retransmission round-trips when possible (FEC).
When loss exceeds FEC capacity, recover the missing frames with a single bitmap NACK exchange instead of full retransmission.
Avoid head-of-line blocking: each voice message stands alone; partial playback is allowed if loss is unrecoverable.
Be codec-agnostic so future codecs (Opus, Codec2, …) can ship without a protocol bump.
Resist abuse: bound per-sender resource use, reject malformed frames early. Confidentiality is delegated to Meshtastic's channel encryption layer (AES-256-CTR with the channel PSK).

Receivers MUST drop any frame whose first byte is not [PROTOCOL_VERSION] (0x03). The version byte exists so future revisions can coexist on the same port without breaking older receivers.

2. Transport

Frames ride on standard Meshtastic data packets:

MeshPacket {
    from:     <sender node num>           // fixed32
    to:       <dest node num | 0xFFFFFFFF for broadcast>
    channel:  <channel index>
    decoded: Data {
        portnum: PRIVATE_APP                // = 256
        payload: <chunk bytes>              // header + body, ≤ 231 B
    }
    want_ack: <true for DMs of DATA/PARITY frames>
}

Field	Value
Port	`PRIVATE_APP` = 256
`to`	Broadcast (`0xFFFFFFFF`) or specific node num
`channel`	Currently selected Meshtastic channel index
`want_ack`	`true` for DM `DATA`/`PARITY`; `false` for broadcasts and `NACK`

MAX_PACKET_SIZE = 231 bytes (Meshtastic LoRa MTU). All frames MUST fit.

2.1 Adaptive pacing

Senders SHOULD wait between successive packets to avoid GATT busy errors and LoRa duty-cycle starvation. The recommended delay depends on the radio's modem preset (read from Config.LoRaConfig.modem_preset):

Modem preset	Pacing
`SHORT_TURBO`	150 ms
`SHORT_FAST`	250 ms
`SHORT_SLOW`	400 ms
`MEDIUM_FAST`	500 ms
`MEDIUM_SLOW`	700 ms
`LONG_FAST`	900 ms
`LONG_MODERATE`	1200 ms
`LONG_SLOW`	1800 ms
`VERY_LONG_SLOW`	3000 ms

When the preset is unknown, senders MUST default to 500 ms. Each value is tuned so pacing >= air_time(recommended_chunk_size) with ~30 % headroom; sending faster overruns the radio's small outbound queue and drops frames before they reach the air.

Local transports (USB-serial, BLE-to-radio) are not subject to LoRa duty-cycle limits, but the radio's queue still benefits from pacing; the recommended values above SHOULD be used regardless of the link to the radio.

2.2 Firmware-queue backpressure

In addition to the time-based pacing above, senders SHOULD honour the firmware's outbound queue depth, which Meshtastic devices advertise via FromRadio.QueueStatus { res, free, maxlen, mesh_packet_id } after every accept/drain. When free drops to a small low-water mark (the reference implementation uses 2), the sender MUST pause until the next QueueStatus update before pushing another voice frame, with a safety timeout (≈ 2 s) so a missed update can't stall transmission indefinitely.

Without this gate, a long voice burst can overflow the firmware's outbound queue and trigger an out-of-memory reboot on the sender device. NACK-driven retransmits MUST flow through the same paced + backpressured path as the initial DATA/PARITY frames.

3. Frame format

Every frame is at most 231 bytes and starts with a 16-byte header (12 logical bytes followed by a 4-byte MAC trailer):

 Byte  Size  Field            Encoding
 ───────────────────────────────────────────────────────────────────────
   0     1   version          UInt8 = 0x03
   1     1   type_flags       UInt8: bits 6-7 packet_type
                                     bit  5  reserved (must be 0; was V2 `encrypted`)
                                     bit  4  last_in_stream
                                     bit  3  reserved (must be 0; was V2 `mac_keyed`)
                                     bits 0-2 reserved (must be 0)
   2     4   message_id       UInt32, big-endian
   6     1   codec            UInt8 (see §3.2)
   7     1   codec_param      UInt8 (codec-specific, see §3.2)
   8     1   stream_seq       UInt8 (per-(from,channel) monotonic)
   9     1   chunk_index      UInt8
  10     1   total_data       UInt8 (1..=255, original data chunks; 0 is reserved and rejected — see §9.2)
  11     1   parity_count     UInt8 (0..=128, FEC parity chunks; 0 = no FEC, the default)
  12     4   mac_tag          SHA-256(header[0..12])[..4] — unkeyed integrity tag
 ───────────────────────────────────────────────────────────────────────
  16   ≤215  body             plaintext codec/parity/NACK bytes (see §3.3)

3.1 Packet types

packet_type is the top 2 bits of type_flags:

Value	Name	`body` content
0	`DATA`	encoded audio for `chunk_index` (0..=total_data−1)
1	`PARITY`	Reed-Solomon parity for `chunk_index` (0..=parity_count−1)
2	`NACK`	bitmap of missing data chunk indices (see §3.4)
3	reserved	receivers MUST drop

3.2 Codec field

`codec`	Name	`codec_param` meaning
0	`AMR_NB`	AMR-NB bitrate ordinal (0..=7), see §3.2.1
1	`OPUS`	bitrate / 1000 (kbps); typical range 6..=64, advisory only
2	`PCM_S16LE`	sample rate index: 0=8 kHz, 1=16 kHz
3	`CODEC2`	Codec2 mode ordinal, see §3.2.2
4..255	reserved	receivers MUST drop unknown codecs

codec_param is codec-specific metadata passed through unmodified by the protocol; receivers SHOULD interpret it per the codec column above but the protocol does not range-check it.

The codec/codec_param fields are advisory: the protocol does not transcode. Receivers that do not support the advertised codec MUST drop the frame and SHOULD surface a "codec unsupported" event to the application layer. codec values in the 4..=255 range are reserved; receivers MUST drop frames carrying them.

3.2.1 AMR-NB bitrates

Ordinal	Mode	kbps	Frame bytes (incl. ToC)
0	MR475	4.75	13
1	MR515	5.15	14
2	MR59	5.90	16
3	MR67	6.70	18
4	MR74	7.40	20
5	MR795	7.95	21
6	MR102	10.2	27
7	MR122	12.2	32

Default: MR795. Frame duration: 20 ms. The AMR file header #!AMR\n is not carried on the wire; senders strip it before chunking and receivers re-prepend it before writing files. The protocol body holds raw codec frames only.

3.2.2 Codec2 modes

Ordinal	Mode	bitrate	Frame bytes
0	3200	3.2 kbps	8
1	2400	2.4 kbps	6
2	1600	1.6 kbps	8
3	1400	1.4 kbps	7
4	1300	1.3 kbps	7
5	1200	1.2 kbps	6

Default: 1200. Frame duration is mode-dependent (20–40 ms). Like AMR-NB, the protocol body carries raw codec frames only; no container header.

3.3 Body layout

The body is always plaintext at this layer; confidentiality on the air comes from Meshtastic's channel encryption wrapping the whole Data payload (header + body) at the firmware level.

DATA frames carry the codec frame bytes for this chunk.
PARITY frames carry the Reed-Solomon parity for this chunk index, sized to chunk_size (see §4).
NACK frames carry the bitmap described in §3.4.

3.4 NACK body

A NACK frame carries the standard 16-byte header, with these constraints:

packet_type = NACK (binary 10 in type_flags bits 6–7)
chunk_index = 0
message_id, codec, codec_param, stream_seq, total_data MUST echo the values of the message being NACK'd. parity_count MUST be echoed for completeness; receivers of NACKs (i.e. the original sender) MAY ignore it since the bitmap is sized solely from total_data.

The body is:

 Byte   Size              Field         Encoding
 ─────────────────────────────────────────────────────────────────────
   0       1              nack_version  UInt8 = 0x01
   1       1              flags         UInt8: bit 0 give_up
                                                bits 1-7 reserved
   2  ⌈total_data/8⌉      bitmap        big-endian, bit `i` set ⇒ index `i` missing

Bit 0 (most-significant bit of byte 2) corresponds to chunk index 0; bit 7 to chunk index 7; byte 3 bit 0 to chunk index 8; and so on. A NACK with all bitmap bits cleared is a positive ACK and the sender SHOULD stop transmitting any remaining queued parity chunks for this message; receivers in the reference implementation rely on natural completion (and the per-message blacklist for late frames) and do not currently emit this shape, but parsers MUST accept it. When give_up is set, the receiver has timed out and the sender SHOULD discard any remaining queued chunks for this message.

NACKs are not retransmitted (the loss-recovery loop is itself the retransmission), and SHOULD be sent with want_ack=false.

Broadcast suppression. Receivers MUST NOT emit NACKs for messages addressed to the broadcast address. With multiple listeners on the same channel, every receiver NACK'ing the same missing chunks would flood the sender, and the sender has no way to choose a retransmit target. Broadcast voice relies on FEC (parity_count) and the partial-on-timeout fallback. The reference receiver short-circuits the NACK emission branch when to == broadcast.

NACK backoff is host-tunable. The reference receiver schedules NACK rounds at nack_window × backoff_base^min(round, 4). The base is selected by the voice.nack_mode setting: typically 2 for short/medium presets (doubling — faster retries), 3 for long-range presets (tripling — gentler). backoff_base = 0 disables NACK emission entirely (used by voice.nack_mode = off and by the broadcast short-circuit above).

4. Chunk size

Variable chunk size per message: each message picks a chunk_size ∈ [16, 219] based on the modem preset:

Modem preset class	`chunk_size`
Short-range (high SNR margin)	219
Medium-range	160
Long-range	96
Very long-range (worst loss profile)	48

chunk_size is not carried in the header — receivers infer it from the first frame whose body length is unambiguous: any PARITY frame, or any non-final DATA frame. A receiver MUST NOT freeze chunk_size from a lone trimmed final DATA chunk that arrives first; it defers discovery until one of the unambiguous frame types arrives. Once established, later DATA frames whose body length differs (excluding the last data chunk, which MAY be shorter) and PARITY frames whose body length differs MUST be rejected.

The sender's recording duration limit derives from chunk_size:

max_audio_bytes(chunk_size) = chunk_size × 255  # 255 = max total_data

At chunk_size = 219 and OPUS @ 16 kbps that's ~28 s of audio per message; at chunk_size = 48 and AMR-NB @ MR795 it's ~12 s.

5. Forward Error Correction

FEC uses Reed-Solomon over GF(2⁸) (reed-solomon-erasure crate) with (total_data, parity_count) shards.

The sender encodes total_data original chunks (zero-padded to chunk_size) and produces parity_count parity chunks. Padding is removed only on the final data chunk during reassembly.
The receiver MUST be able to reconstruct the message if it has any total_data shards out of the total_data + parity_count total (any combination of DATA and PARITY shards counts toward the threshold). Loss up to parity_count shards is recoverable without a NACK.
Final-chunk caveat. The trimmed length of the final DATA chunk is not carried on the wire (§3.3); receivers only learn it when they observe the real final DATA frame. If the final DATA chunk is the one missing, FEC alone cannot recover it without inventing trailing zero-padding. Receivers MUST NOT finalize a message via FEC when the final DATA shard's real length is unknown — they MUST either wait for the real frame (via NACK-driven retransmit) or fall back to a partial finalize on hard timeout. The reference implementation defers FEC recovery of the last shard in exactly this case.
parity_count is sender policy; recommended values:

Mesh profile	`parity_count` (% of `total_data`)
Short / quiet	10 %
Medium / mixed	20 %
Long / lossy	33 %
Broadcast (no NACK feedback channel)	50 %

parity_count = 0 is allowed and disables FEC entirely (NACK still works).

When loss exceeds parity_count, the receiver issues a NACK; the sender retransmits the missing data chunks (only) and MAY add additional parity chunks beyond the original parity_count. Retransmitted frames carry the same message_id, chunk_index, and total_data. parity_count MAY grow on retransmit; receivers MUST accept frames whose parity_count is ≥ the value first observed and SHOULD reject decreases. Retransmitted DATA / PARITY frames are paced under the same rules as the original send (see §2.1).

6. Identification keys

Message identity: (from_node_num, message_id) where message_id is a non-zero u32 chosen by the sender. The 32-bit space makes accidental collision with the receiver's recently-finalized blacklist negligible. from MUST be the lowercase 8-hex-digit form !hex8; ids should be normalized before lookup.
Stream identity: (from_node_num, channel, stream_seq) where stream_seq is a u8 monotonic counter per (sender, channel) pair, wrapping at 256. It is intended for receivers that order overlapping voice messages from the same sender on the same channel deterministically (e.g. interleaved recordings); the reference implementation currently treats it as informational and echoes it on NACK frames. The last_in_stream flag (bit 4 of type_flags) is set on the final frame of a recording session; receivers MAY use it to expire stream-history state for that sender. The reference implementation currently treats this bit as informational.

7. Confidentiality and integrity

Confidentiality is provided by Meshtastic's channel encryption layer: the whole Data payload (this protocol's header + body) is encrypted with AES-256-CTR using the channel PSK before being handed to the LoRa PHY. Peers without the channel PSK cannot read the bytes.

This protocol layer does not add a per-message AEAD envelope. V2 carried one (AES-256-GCM with an HKDF-derived per-message key); it was removed in V3 because Meshtastic's channel encryption already covers the threat we care about (an outsider passively snooping the air) and the GCM nonce + tag burned 28 octets of every chunk — meaningful on LoRa where the body budget is ~199 bytes.

The 4-byte trailing mac_tag is SHA-256(header[0..12])[..4], unkeyed. It serves two purposes:

Bit-flip detection. AES-CTR is malleable: an attacker who can modify ciphertext bits on the air can flip the corresponding bits of the underlying plaintext header (chunk_index, parity_count, message_id, etc.). The SHA-256 truncated tag catches such tampering before any field is acted on.
Accidental-corruption check. On marginal links, occasional in-frame bit errors slip past LoRa FEC; the tag rejects them rather than silently corrupting the reassembly state.

The tag is not an authenticator: any peer who holds the channel PSK (and can therefore both encrypt new frames and read existing ones) can forge a valid mac_tag. Authentication of the audio stream against malicious channel members is out of scope for this layer.

8. Send-side flow

                     ┌─────────────────┐
                     │ codec encoder   │ produces packed audio bytes
                     └────────┬────────┘
                              │
                              ▼
                     ┌─────────────────┐
                     │ split into      │ chunk_size from §4
                     │ total_data      │
                     │ chunks (last    │
                     │ may be padded)  │
                     └────────┬────────┘
                              │
                              ▼
                     ┌─────────────────┐
                     │ Reed-Solomon    │ produce parity_count parity shards
                     │ encode          │
                     └────────┬────────┘
                              │
                              ▼
                     ┌─────────────────┐
                     │ for each shard: │ build header, send via PRIVATE_APP
                     │   send_data     │ with adaptive pacing
                     │   want_ack=true │
                     │   for DM        │
                     └────────┬────────┘
                              │
                              ▼
                     ┌─────────────────┐
                     │ wait for NACK   │ up to nack_window_ms
                     │ window          │
                     └────────┬────────┘
                              │
                              ▼
                     ┌─────────────────┐
                     │ on NACK:        │ retransmit missing data chunks
                     │ rebuild missing │ (and optionally extra parity)
                     │ shards          │
                     └─────────────────┘

Senders MAY also drop silence chunks: a DATA chunk whose payload is entirely codec NO_DATA frames does not need to be sent, since the receiver synthesises silence for missing chunks anyway. Silence detection is codec-specific (e.g. AMR-NB: all bytes equal to 0x7C). Receivers MUST NOT distinguish silence-skipped chunks from lost chunks: both appear as "missing", are eligible for FEC reconstruction, and are zero-filled on timeout per §9.

The full send is cancellable: a CancellationToken signal aborts remaining transmissions and emits last_in_stream = 1 on the next sent frame.

9. Receive-side flow

chunk arrives
      │
      ▼
parse 16-byte header  ─────► reject if version != 3
                     ─────► reject if any reserved type_flags bit is set
                     ─────► reject if header MAC mismatches (§3)
      │
      ▼
check blacklist        ─────► reject if already finalized
      │
      ▼
lookup or create AssemblyState for (from, message_id)
      │
      ├─ NACK frame? ─────► route to send-side state for that message_id
      │
      ├─ new state → start internal timer (chunk_timeout_seconds)
      │              and per-sender rate-limit slot
      │
      ▼
store body at chunks[chunk_index] (DATA) or parity[chunk_index] (PARITY)
      │
      ├─ enough shards to RS-decode? ──► reconstruct missing data chunks
      │
      ├─ all data chunks present? ─────► finalize (complete) → emit VoiceMessage
      │
      ├─ partial timeout?           ─────► emit NACK with bitmap of missing
      │                                    chunks; reset timer
      │
      └─ hard timeout (after N NACK rounds)? ─► finalize (partial) or discard
                                                per partial_play_on_timeout

9.1 Resource bounds

MAX_IN_PROGRESS_GLOBAL = 64 total reassemblies. When the global cap is reached and a new (from, message_id) arrives, the receiver evicts the in-progress entry with the oldest started_at and blacklists its key for BLACKLIST_TTL.
MAX_IN_PROGRESS_PER_SENDER = 4 per from_node_num (prevents one chatty peer from starving everyone else).
MAX_MESSAGE_BYTES = 255 × 219 = 55_845 (worst-case sum of data shards before FEC overhead). Refused beyond this.
BLACKLIST_TTL = 600 s, BLACKLIST_MAX = 100. This is the receiver's completion-memory window: once a (from, message_id) pair has finalized (complete or partial), late chunks for that pair are silently dropped for this long so the sender's firmware-queue drain (which can outrun the receiver's completion by tens of seconds on slow presets) doesn't resurrect a phantom partial reassembly. The window should be ≥ the assembler's message_timeout.
NACK_MAX_ROUNDS = 400 consecutive rounds without progress before the receiver gives up. The counter resets every time a new shard lands, so a sender that's still servicing every NACK round keeps the assembly slot alive (capped only by message_timeout); only 400 rounds in a row with zero new chunks finalizes partial. The 400 is the default-config seed: hosts that change message_timeout or nack_window re-derive it as ⌈message_timeout / nack_window⌉, so the per-message timeout (default 1200 s; 1200 s / 3 s = 400) is the only practical ceiling. Earlier revisions used a cumulative counter (never reset) at 3 (gave up after ~4-5 s) and 32 (~48 s); both surfaced phantom partial: N/M finalizes on healthy slow-trickle messages where inter-chunk gaps routinely exceed a second.
NACK_WINDOW_MS = 3000 after the last seen chunk before issuing a NACK. Receivers apply host-tunable exponential backoff between rounds (nack_window × backoff_base^min(round, 4)); the long-range default base of 3 gives 3 s, 9 s, 27 s, 81 s, 243 s (cap), so a stuck sender is probed aggressively at first but does not flood the channel indefinitely. See §3.4 / the voice.nack_mode setting for the 2/3 base selection. Global NACK emission is bounded by MAX_IN_PROGRESS_GLOBAL per tick; senders / transports SHOULD pace NACK transmission per §2.1.
dead_sender_timeout = 120 s (experimental flood-control). If no DATA or PARITY frame arrives for this long the sender is presumed dead and NACKs are suppressed until message_timeout, so the receiver stops firing cap-multiplied NACKs at a peer that has left the mesh. Validated to be < message_timeout.
MAX_VALIDATION_STRIKES = 3 (experimental resource-bounding). After this many post-template validation failures (codec / total_data / stream_seq mismatch) on the same in-progress entry, the entry is evicted and blacklisted to free its scarce per-sender slot from a chatty bad sender.

9.2 Rejection rules

A receiver MUST reject (silently drop) a frame when any of the following hold:

version != 0x03 (protocol version 3)
any reserved bit of type_flags is set (0x2F mask — covers the former V2 encrypted (0x20) and mac_keyed (0x08) bits plus the three low reserved bits)
packet_type == 3 (reserved)
total_data == 0
chunk_index >= total_data for DATA frames
chunk_index >= parity_count for PARITY frames
parity_count > 128
the frame's codec is unknown to the receiver (spec §3.2)
the frame's codec differs from the codec established by the first frame of this (from, message_id)
the frame's total_data differs from the value established by the first frame of this (from, message_id)
the frame's parity_count is less than the value established by the first frame of this (from, message_id) (spec §5: it MAY grow on retransmit but MUST NOT shrink)
DATA body length differs from the message's established chunk_size and this is not the last data chunk
PARITY body length differs from the message's established chunk_size
a NACK frame's chunk_index is non-zero (spec §3.4)
the (from, message_id) is on the recently-completed blacklist
the sender already has MAX_IN_PROGRESS_PER_SENDER in-flight messages and this is a new message_id

10. Data model (Rust)

pub struct VoiceMessage {
    pub message_id: u32,
    pub from: String,                       // "!aabbccdd"
    pub to: VoiceDestination,               // Node(u32) | Broadcast
    pub stream_seq: u8,
    pub codec: VoiceCodec,
    pub codec_param: u8,
    pub audio: Vec<u8>,                     // codec frame bytes, no container
    pub timestamp: chrono::DateTime<chrono::Utc>,
    pub is_complete: bool,
    pub total_data: u8,
    pub received_data: u8,
    pub recovered_via_fec: u8,              // chunks reconstructed by RS
    pub channel: u32,
}

pub enum VoiceCodec { AmrNb, Opus, PcmS16Le, Codec2, Unknown(u8) }
pub enum VoiceDestination { Node(u32), Broadcast }

The reassembled audio is the raw codec stream; the protocol does not prepend any container header. Callers wrap the bytes in the appropriate container themselves if needed (e.g. #!AMR\n for AMR-NB playback).

11. Limitations and design trade-offs

Constraint	Value	Reason
Max chunks per message	255 (`u8`)	header byte
Max parity per message	128	RS coder limit
Max audio per message	55 845 B	255 × 219
Min chunk size	16 B	per-frame overhead floor
Stream sequence wrap	256	per-(from,channel)
Confidentiality	delegated to Meshtastic channel AES-CTR	no per-message envelope at this layer
Header integrity	SHA-256[..4] (unkeyed)	catches bit-flips against AES-CTR malleability
FEC	RS over GF(2⁸)	survives any `parity_count` losses
Max NACK rounds	400 cumulative	bounds total airtime per message
Codec	application-decided	protocol carries opaque bytes

The protocol explicitly does not provide:

a built-in audio codec (out of scope; bring-your-own)
end-to-end authentication or confidentiality beyond what Meshtastic's channel encryption already gives you. Any peer with the channel PSK can forge any frame on this layer (DATA, PARITY, NACK including give_up). A future revision MAY reintroduce a per-message AEAD envelope or a keyed header MAC if a concrete threat model warrants it — see git history (V2) for the previous design.
congestion control (relies on adaptive pacing + per-sender rate limit)
ordering across messages (only within a stream via stream_seq)

Appendix A: Reference constants

pub const PROTOCOL_VERSION: u8 = 0x03;
pub const HEADER_SIZE: usize = 16;
pub const MAX_PACKET_SIZE: usize = 231;
pub const MAX_BODY_SIZE: usize = MAX_PACKET_SIZE - HEADER_SIZE;  // 215
pub const MAX_MESSAGE_BYTES: usize = MAX_CHUNKS_PER_MESSAGE * MAX_BODY_SIZE;  // 55_845
pub const MIN_CHUNK_SIZE: usize = 16;
pub const MAX_CHUNKS_PER_MESSAGE: usize = 255;
pub const MAX_PARITY_PER_MESSAGE: usize = 128;
pub const MAX_IN_PROGRESS_GLOBAL: usize = 64;
pub const MAX_IN_PROGRESS_PER_SENDER: usize = 4;
pub const BLACKLIST_TTL: Duration = Duration::from_secs(600);
pub const BLACKLIST_MAX: usize = 100;
pub const NACK_MAX_ROUNDS: u16 = 400;
pub const NACK_WINDOW_MS: u64 = 3000;

Appendix B: Type/flag bit layout

   type_flags byte:
     bit 7  bit 6  bit 5      bit 4           bit 3      bits 2..0
     ┌──────────────┐ ┌────────┐ ┌──────────────┐ ┌────────┐ ┌──────────┐
     │ packet_type  │ │reserved│ │last_in_stream│ │reserved│ │ reserved │
     └──────────────┘ └────────┘ └──────────────┘ └────────┘ └──────────┘
       (2 bits)       (=0; V2    (1 bit)         (=0; V2    (3 bits, =0)
                      encrypted)                  mac_keyed)

Field	Mask
`packet_type`	`0xC0`
reserved (V2 `encrypted`)	`0x20`
`last_in_stream`	`0x10`
reserved (V2 `mac_keyed`)	`0x08`
reserved	`0x07`

Uh oh!

Voice Protocol

Voicetastic Voice Protocol

1. Goals

2. Transport

2.1 Adaptive pacing

2.2 Firmware-queue backpressure

3. Frame format

3.1 Packet types

3.2 Codec field

3.2.1 AMR-NB bitrates

3.2.2 Codec2 modes

3.3 Body layout

3.4 NACK body

4. Chunk size

5. Forward Error Correction

6. Identification keys

7. Confidentiality and integrity

8. Send-side flow

9. Receive-side flow

9.1 Resource bounds

9.2 Rejection rules

10. Data model (Rust)

11. Limitations and design trade-offs

Appendix A: Reference constants

Appendix B: Type/flag bit layout

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Clone this wiki locally