voiceagent-esp32

English | 简体中文

ESP32 Rust SDK for the Streamcore Voice Agent server. API design mirrors LiveKit's client-sdk-esp32: create an Agent handle once, connect/disconnect at will, drive everything from your own main loop. Built on Espressif's native esp_peer C library — runs on ESP-IDF / FreeRTOS, no tokio, no full OS networking required.

This crate is a library, not firmware. It has no main() and produces no flashable image. Runnable firmware lives in examples/esp32 — start there if you want something on a board today, then come back here when you're writing your own app.

---

Mental model

Three things to know:

1. You own the media pipeline. The SDK ships a Capturer trait + I2S default impl, and a Renderer trait + I2S default impl. Bring your own implementations for offline capture, USB mics, custom DACs — the SDK only cares that PCM moves in and out.
2. You own the event loop. Agent::connect is non-blocking; the agent runs in its own thread. Your main() is free to drive buttons, LEDs, sleep modes, reconnect logic — anything.
3. Server interaction is callback + RPC. Built-in callbacks for state changes, transcripts, AI responses, errors, and topic-addressed data packets. Custom device tools (camera, sensors, GPIO) are exposed via Agent::rpc_register so the AI can call them by name.

use voiceagent_esp32 as va;

fn main() -> anyhow::Result<()> {
    va::system_init();
    let p = esp_idf_svc::hal::peripherals::Peripherals::take()?;

    // 1. WiFi (the SDK no longer does this for you).
    let _wifi = va::wifi::connect_sta(p.modem, env!("WIFI_SSID"), env!("WIFI_PASSWORD"))?;

    // 2. Build the media pipeline.
    let mic = va::I2sMicCapturer::new(p.i2s1, /* bclk */, /* din */, /* ws */, true)?;
    let speaker = va::I2sSpeakerRenderer::new(p.i2s0, /* bclk */, /* dout */, /* ws */, 24_000)?;

    // 3. Create the agent (no network activity yet).
    let agent = va::Agent::create(va::AgentOptions {
        publish:   Some(va::PublishOptions { capturer: Box::new(mic) }),
        subscribe: Some(va::SubscribeOptions { renderer: Box::new(speaker) }),
        on_state_changed: Some(Box::new(|s| log::info!("state: {s:?}"))),
        on_transcript:    Some(Box::new(|t, f| log::info!("user: {t} ({f})"))),
        on_response:      Some(Box::new(|t|    log::info!("ai: {t}"))),
        ..Default::default()
    })?;

    // 4. Expose a tool the AI can call.
    agent.rpc_register("get_temperature", |inv| {
        inv.return_ok(serde_json::json!({"celsius": 22.4}));
    })?;

    // 5. Connect — returns immediately.
    agent.connect(env!("WHIP_ENDPOINT"), None)?;

    // 6. Your main loop drives buttons, sleep, reconnect, etc.
    loop { std::thread::sleep(std::time::Duration::from_secs(1)); }
}

---

Run an example first

git clone https://github.com/streamcoreai/examples.git
cd examples/esp32
cp .env.example .env        # WiFi + WHIP endpoint
cargo build --release && espflash flash --monitor \
  target/xtensa-esp32s3-espidf/release/voice_agent

Full prerequisites, board configuration, and troubleshooting: examples/esp32/README.md.

Three targets ship there:

• voice_agent — WiFi + mic + speaker + display + camera + push-to-talk + RPC.
• minimal_audio — WiFi + mic + speaker, ~80 lines total.
• headless — no display/camera; demonstrates publish_data, on_data, custom RPC, and topic-based heartbeat.

---

Using the SDK in your own project

An ESP-IDF binary crate needs more scaffolding than a normal Rust one. The fastest route is to copy examples/esp32 and delete what you don't need. If you'd rather start clean, here's the complete list.

1. Toolchain

cargo install espup ldproxy espflash
espup install
. $HOME/export-esp.sh        # every shell — exports LIBCLANG_PATH

2. Dependency

[dependencies]
voiceagent-esp32 = "0.1"

# Your crate is the binary, so it — not the SDK — needs `binstart`.
esp-idf-svc = { version = "0.52", features = ["binstart", "critical-section"] }
esp-idf-hal = "0.46"
esp-idf-sys = { version = "0.37", features = ["binstart"] }
anyhow = "1"
log = "0.4"
serde_json = { version = "1", default-features = false, features = ["alloc"] }

[build-dependencies]
embuild = "0.33"

Working from a checkout of this repo instead: voiceagent-esp32 = { path = "../esp32" }.

esp_peer comes from a git submodule inside this crate, contributed to your build automatically through [package.metadata.esp-idf-sys] extra_components. Initialise it once in the SDK checkout, or the ESP-IDF component resolution step fails:

git -C path/to/esp32 submodule update --init --recursive

3. build.rs

fn main() {
    embuild::espidf::sysenv::output();
}

4. .cargo/config.toml

[build]
target = "xtensa-esp32s3-espidf"

[target.xtensa-esp32s3-espidf]
linker = "ldproxy"
runner = "espflash flash --monitor"     # optional: makes `cargo run` flash
rustflags = ["--cfg", "espidf_time64"]

[unstable]
build-std = ["std", "panic_abort"]

[env]
ESP_IDF_VERSION = "v5.4"
ESP_IDF_COMPONENT_MANAGER_ENABLED = "1"
MCU = "esp32s3"

5. rust-toolchain.toml

[toolchain]
channel = "esp"

6. idf_component.yml

The SDK links against three managed ESP-IDF components. They resolve against the root (binary) crate, so this manifest belongs in your project:

dependencies:
  espressif/esp_audio_codec: { version: "~2.3.0" }   # Opus encode/decode
  espressif/esp-sr:          { version: "^1.9.0" }   # AFE — AGC + noise suppression
  espressif/esp32-camera:    { version: "^2.0.0" }   # only if you use va::camera
  idf:                       { version: ">=5.1.0" }

7. sdkconfig.defaults

Copy sdkconfig.defaults verbatim to start. The settings that are not optional:

Setting	Why
CONFIG_SPIRAM=y (+ mode/speed)	Opus, AFE, and DTLS buffers do not fit in internal RAM
CONFIG_MBEDTLS_SSL_PROTO_DTLS=y, CONFIG_MBEDTLS_SSL_DTLS_SRTP=y, CONFIG_MBEDTLS_X509_CREATE_C=y	esp_peer needs DTLS-SRTP and generates a self-signed cert
CONFIG_MBEDTLS_DEFAULT_MEM_ALLOC=y	Lets mbedTLS spill to PSRAM; internal-only allocation fails once esp-sr has taken its DRAM
CONFIG_ESP_MAIN_TASK_STACK_SIZE=65536	Agent setup runs deep on the main task
CONFIG_AFE_INTERFACE_V1=y, CONFIG_SR_NSN_WEBRTC=y, CONFIG_SR_VADN_WEBRTC=y	AFE pipeline used by I2sMicCapturer

8. partitions.csv

The app image is a few MB with ESP-IDF, mbedTLS, and the esp-sr models linked in. A 3 MB factory partition on ≥ 4 MB flash works:

# Name,   Type, SubType, Offset,  Size,   Flags
nvs,      data, nvs,     ,        0x6000,
phy_init, data, phy,     ,        0x1000,
factory,  app,  factory, ,        0x300000,

---

API reference

Agent

Agent is Clone (an Arc inside) and Send + Sync — clone it into callbacks and other threads freely.

Method	Behaviour
Agent::create(AgentOptions) -> Result<Agent>	Spawns the worker thread. No network activity yet.
connect(whip_endpoint: &str, token: Option<&str>) -> Result<()>	Non-blocking. token is a bearer token, sent as Authorization: Bearer …. Progress arrives on on_state_changed.
disconnect() -> Result<()>	Tears down the peer and sends the WHIP DELETE.
state() -> ConnectionState	Disconnected | Connecting | Connected | Failed.
set_mic_enabled(bool) -> Result<()>	Push-to-talk. Calls Capturer::set_enabled on the worker thread.
publish_data(topic: &str, data: &[u8]) -> Result<()>	Topic-addressed packet over the data channel.
rpc_register(method, handler) -> Result<()>	Handler is FnMut(RpcInvocation) + Send + 'static.
rpc_unregister(method) -> Result<()>	Removes a handler.

Need a JWT rather than a static key? whip::fetch_token(token_url, api_key) POSTs to your server's token endpoint and returns the token field, ready to hand to connect.

AgentOptions

Every field is optional; ..Default::default() covers the rest.

Field	Type	Purpose
publish	Option<PublishOptions>	Outbound audio. Omit for subscribe-only devices.
subscribe	Option<SubscribeOptions>	Inbound audio. Omit for publish-only devices.
stun_server	Option<String>	e.g. "stun:stun.l.google.com:19302". Not needed on a LAN.
on_state_changed	FnMut(ConnectionState)	Connection lifecycle.
on_transcript	FnMut(&str, bool)	User speech; the bool is is_final.
on_response	FnMut(&str)	AI text, chunk by chunk.
on_error	FnMut(&str)	Server-reported error.
on_data	FnMut(&str, &[u8])	Inbound topic + payload.
on_raw_event	FnMut(&str)	Any data-channel JSON the SDK didn't handle.
on_wake_word	FnMut()	On-device wake word; requires a WakeNet capturer (below).

Callbacks run on the agent worker thread. Keep them short — blocking here stalls the audio path.

Capture and render

// 16 kHz mono I2S mic + ESP-SR AFE (AGC + noise suppression).
let mic = va::I2sMicCapturer::new(i2s1, bclk, din, ws, /* use_afe */ true)?;

// Same, plus on-device wake-word detection. The model comes from sdkconfig
// (CONFIG_SR_WN9_HIESP=y etc.); detections fire `on_wake_word`.
let mic = va::I2sMicCapturer::new_with_wakenet(i2s1, bclk, din, ws, true, true)?;

// I2S speaker. Output rate must be 16_000 or 24_000.
let speaker = va::I2sSpeakerRenderer::new(i2s0, bclk, dout, ws, 24_000)?;

Both are trait objects, so swapping in your own hardware is two methods:

pub trait Capturer: Send {
    fn read_frame(&mut self) -> Option<Vec<i16>>;   // 16 kHz mono i16, ~320 samples
    fn set_enabled(&mut self, enabled: bool) -> Result<()>;
    fn consume_wake_event(&mut self) -> bool { false }
}

pub trait Renderer: Send {
    fn render_audio(&mut self, pcm: &[i16]) -> Result<i32>;   // 16 kHz mono i16
}

NullCapturer and NullRenderer are there for one-way devices and bring-up.

RPC

agent.rpc_register("set_led", |inv| {
    // inv.id, inv.method, inv.params (serde_json::Value)
    match inv.params.get("on").and_then(|v| v.as_bool()) {
        Some(on) => { drive_led(on); inv.return_ok(serde_json::json!({"ok": true})); }
        None     => inv.return_err(400, "missing 'on'"),
    }
})?;

Exactly one of return_ok / return_err per invocation — the server is waiting on a matching rpc.response.

---

Capabilities

Stage	Implementation
WiFi STA	esp-idf-svc::wifi (helper: va::wifi::connect_sta)
WHIP signaling	ESP-IDF HTTP client (esp_http_client)
WebRTC stack	esp_peer (ICE + DTLS + SRTP + SCTP data channel)
Mic capture	I2sMicCapturer — I2S RX 16 kHz mono + ESP-SR AFE
Speaker output	I2sSpeakerRenderer — I2S TX 16 kHz or 24 kHz
Audio codec	Opus 16 kHz @ 32 kbps (via esp_audio_codec)
Wake word	Optional ESP-SR WakeNet via `new_with_wakenet$
\text{Display}	\text{Optional} \text{ST7789} 240 \times 280 \text{helper} ($va::display::DisplayHandle`)
Camera	Optional OV2640 JPEG capture (va::camera)
RPC	Topic-based remote method calls + structured invocations
Data	Topic-addressed packets via publish_data / on_data

---

Wire protocol (data channel)

The SDK speaks a small JSON protocol over the WebRTC data channel. Anything not in this list is delivered verbatim to your on_raw_event callback so you can extend the protocol freely.

type	Direction	Purpose
transcript	inbound	User speech, partial or final
response	inbound	AI text response chunk
error	inbound	Server-reported error
rpc.request	inbound	Server invokes a registered RPC method
rpc.response	outbound	Device answers an rpc.request
data	both	Application-defined topic + payload

---

Building blocks (advanced)

Every layer is exposed as a pub module so you can compose your own pipeline:

Module	Contents
agent	Agent, AgentOptions, ConnectionState
capture	Capturer, I2sMicCapturer, NullCapturer
render	Renderer, I2sSpeakerRenderer, NullRenderer
rpc	RpcInvocation
display	ST7789 thread + DisplayHandle
camera	OV2640 init + base64 chunked send
wifi	Standalone WiFi STA helper
whip	WHIP HTTP client + fetch_token
webrtc	Safe wrapper around esp_peer (raw)
opus_codec	Opus encode/decode
afe_pipeline	ESP-SR AFE pipeline (AGC + NS)
audio	Low-level I2S mic/speaker drivers

---

Repository layout

esp32/                          this crate — the SDK library
├── src/                        agent, capture, render, rpc, whip, webrtc, …
├── components/
│   ├── esp-webrtc-solution/    git submodule — provides esp_peer
│   └── voiceagent_vc_frontend/ ESP-SR AFE wrapper component
└── sdkconfig.defaults          reference config to copy into your app

examples/esp32/                 firmware that uses this crate
└── src/bin/{voice_agent,minimal_audio,headless}.rs

cargo build here type-checks the library. It does not produce a flashable image — that's examples/esp32's job.

---

License

MIT — see LICENSE.