A SIP voice endpoint on an ESP32-S3 — registers to a PBX, places a call, and carries two-way G.711 audio peer-to-peer. A working answer to the long-standing "SIP voice on ESP32" ask in ESPHome feature-request #3112.

Status: validated proof-of-concept. The full path — REGISTER → INVITE → ring → answer → bidirectional RTP audio → BYE → clean teardown — has been confirmed on real silicon (LilyGO T3-S3 MVSRBoard) against a live pocket-dial server. It is not production firmware: half-duplex, no jitter buffer, no digest auth (see Constraints). It is a clean, honest base to build on.

pocket-dial (and its upstream, BarGabriel/SipServer) is a SIP registrar/proxy — it brokers signaling but never touches audio. The #3112 ask is the other half: a SIP endpoint that actually moves voice. tincan is that endpoint. It speaks just enough SIP to register and set up a call, then streams G.711/RTP directly to the far end (the proxy stays out of the media path), so a $10 MCU coordinates and carries a real phone call.

        ESP32-S3 (tincan)                      pocket-dial PBX            ext 102 (phone)
   mic ─I2S─► G.711 ─► RTP ─┐                  192.168.x.2:5060            192.168.x.181
                            ├─ SIP REGISTER/INVITE/ACK/BYE ─►  ──routes──►  rings / answers
   spk ◄I2S─ G.711 ◄─ RTP ─┘                                                     │
        ▲                                                                        │
        └──────────────── RTP (G.711 µ-law, UDP) flows PEER-TO-PEER ─────────────┘

LilyGO T3-S3 MVSRBoard — board revision V1.1 (silkscreen). ESP32-S3 (dual Xtensa LX7 @ 240 MHz), 4 MB flash, QSPI PSRAM present but unused.

Why the split: on the ESP32-S3, PDM RX is only available on I2S0, so the mic owns I2S0 and the speaker (standard Philips mode) takes I2S1 — two independent controllers. All board specifics live in main/board_mvsr.h; a different board is a single-file swap. (V1.0 boards ship an MSM261 I2S mic on BCLK 47/WS 15/DATA 48 and need i2s_channel_init_std_mode instead — noted in the header.)

Memory budget: firmware image ≈ 830 KB in a 1.5 MB factory partition (SINGLE_APP_LARGE); ≈ 289 KB internal RAM free for heap at runtime. Comfortable headroom.

Single-task design — app_main (main task, CPU0) runs setup, then the media loop. No dynamic allocation in the hot path; SIP message buffers are stack-local and bounded.

boot ─► NVS ─► audio_init (PDM mic + amp, amp OFF) ─► Wi-Fi STA (blocks until IP)
     ─► SIP REGISTER ext 500 (no auth)  ─► INVITE ext 102 via proxy
     ─► [180 Ringing, up to ~120 s]      ─► 200 OK ─► ACK ─► amp ON
     ─► media loop (push-to-talk, 20 ms frames)
     ─► peer BYE detected ─► amp OFF ─► idle

• Audio path: PDM mic captured at 16 kHz through the S3's hardware PDM→PCM filter (SOC_I2S_SUPPORTS_PDM2PCM), decimated 2:1 to the 8 kHz G.711 rate; speaker fed at 8 kHz on the second I2S controller.
• Codec / framing: G.711 µ-law (PCMU, RTP payload type 0), 8 kHz mono, 20 ms / 160-sample frames. The blocking I2S read paces the loop — no software timer needed.
• Half-duplex by design: PTT gates direction, so mic and speaker are never live simultaneously → acoustic echo cancellation is sidestepped entirely, removing the single hardest part of ESP32 two-way voice.
• Amp gating: SD_MODE is held low except during a connected call — eliminates the continuous I2S-underrun click while booting/ringing/idle.
• Clean teardown: the media loop polls the signaling socket (non-blocking) for an inbound BYE; on hangup it stops, mutes the amp, and idles. RTP underrun feeds a silence frame so the DMA never repeats a stale buffer.
• SIP layer: vendored from pocket-dial (MIT) for message (de)serialization; the user-agent client logic (main/sip_uac.cpp) — REGISTER/INVITE/ACK/BYE construction, response parsing, SDP endpoint extraction — is original.

Requires ESP-IDF (validated on v5.2.1; also builds on v6.0.1 — the CMake guards the v6 esp_driver_* component split).

# 1. Configure your network + targets
#    edit main/poc_config.h: POC_WIFI_SSID / POC_WIFI_PASS
#    (defaults: server 192.168.12.2:5060, register as 500, call 102)

# 2. Build + flash + monitor
idf.py set-target esp32s3
idf.py -p <PORT> flash monitor

A peer at ext 102 must be registered to the same pocket-dial server and able to answer. On boot the board auto-dials it; answer, hold BOOT to talk, release to listen, hang up to end.

Confirmed end-to-end on a LilyGO T3-S3 MVSRBoard ↔ pocket-dial server (Raspberry Pi), serial + server-side journalctl cross-checked:

registered as 500
INVITE 102 → 100 Trying → 180 Ringing → 200 OK → ACK
P2P media to 192.168.12.181  (two-way G.711, intelligible both directions)
<- BYE (peer hung up) → call ended → idle    (no looping, no clicking)

The footguns that bit during bring-up — documented so they don't bite again:

• thread_local TLS bloat is lethal on small RTOS tasks. The vendored IDGen originally used thread_local std::mt19937 (~2.5 KB). ESP-IDF copies the TLS template into every task stack, overflowing the idle (~1.5 KB) and IPC (~1.3 KB) tasks at creation → boot crash (esp_ipc_init assert on v6.0.1; SMP-scheduler LoadProhibited on v5.2.1 — same bug, two faces). Fixed by using esp_random() (stateless, zero TLS). Audit vendored C++ for thread_local before putting it on an MCU.
• IDF v6 driver split: driver/i2s_std.h, driver/i2s_pdm.h, driver/gpio.h moved to esp_driver_i2s / esp_driver_gpio on v6.0. CMake selects per IDF_VERSION_MAJOR.
• Flash mode: DIO (the board's flash didn't benefit from QIO; DIO is the safe default here).
• No jitter buffer — direct play-out. Fine on a quiet LAN; will need a small adaptive buffer for congested networks.
• No acoustic echo cancellation — intentional (half-duplex PTT). Required before full-duplex.
• No digest auth — relies on pocket-dial's open registrar. Add MD5 digest (RFC 3261 §22) for a real PBX.

Adaptive jitter buffer · digest auth · full-duplex + AEC · inbound call handling · dial-on-button (vs auto-dial on boot) · packaging as an ESPHome external component.

MIT — see LICENSE. Dual copyright:

• Original work (audio pipeline, G.711, RTP, the sip_uac SIP client, Wi-Fi/board integration, application code) — © 2026 GlomarGadaffi.
• Vendored SIP message parser in components/sip_core/ (SipMessage, SipSdpMessage, SipMessageFactory, SipMessageTypes, IDGen) — © 2022 BarGabriel, derived via pocket-dial from BarGabriel/SipServer (MIT).

Also referenced (credited, no obligation): the G.711 CCITT/Sun reference companding algorithm, and LilyGO's T3-S3-MVSRBoard pin_config.h for the pin map. Built on MIT code only — no copyleft dependencies.