3. AI Agent Settings

1. AI Xiaozhi "Brain" Architecture

Before diving into configuration, understand AI Xiaozhi's cloud intelligence architecture—this determines your basis for choosing which AI platform:

2. Register and Configure AI Agent

2.1 Register Account and Add Device

AI Xiaozhi's official server xiaozhi.me has pre-integrated ASR + LLM + TTS complete pipeline; you can register and use it directly for free.

1. Browser visits https://xiaozhi.me, register with phone number

2. After login, go to console → Console

3. Click Add Device, select "Open Source", system generates device activation code

The activation code is the credential for device connection to server; used during device activation in 2. Network and Configuration, please keep it safe.

2.2 Configure Large Language Model

After login, you can change the large language model on the console's Configure page:

The official service uses Qwen real-time model free quota by default. If you have a GitHub account, you can login to unlock more model choices.

2.3 Configure Language and Voice

Supports 30+ languages (Chinese, English, Japanese, Cantonese, etc.), covering multi-language real-time interaction; also provides multiple TTS voice options:

2.4 Configure Personality

On the Personality page, you can customize Xiaozhi's personality, tone, and reply style, making the AI more suited to your use case:

3. Voice Wake Word Configuration

Default wake word: Hi,Andy (English)

The wake word is detected locally offline by the ESP-SR engine and does not require internet. You can switch wake word models and feature switches in menuconfig (refer to 3. Serial Communication and Storage System → 4. PSRAM External Memory Experiment):

1.Basic Wake Solution Selection

Configuration path: Xiaozhi Assistant → Wake Word Implementation Type

Purpose: Switch underlying wake word engine solution (default: Wakenet wake model with AFE audio front-end processing), supports switching Chinese / English / Japanese preset wake models.

2.Enable Standby Wake Detection

Configuration path: Xiaozhi Assistant → Enable Wake Word Detection in Listening Mode

After checking this option, device enters low-power monitoring mode, can offline recognize wake word and be awakened.

3.Complete Custom Offline Wake Word Configuration

1. Enable Custom Mode

Path: Xiaozhi Assistant → Wake Word Implementation Type

Switch to Multinet model (Custom Wake Word) to unlock custom configuration below; if using default option, can only use "Hi,Andy" fixed.

2. Three Wake Word Parameters

Custom Wake Word: Fill in pinyin with spaces, used for underlying speech recognition, Chinese characters cannot be filled

Custom Wake Word Display: For screen and log prompts only, can freely modify Chinese text

Custom Wake Word Threshold (%): 0~100 threshold, lower number means more sensitive and easier to false trigger, higher number means more strict and harder to false trigger

3. Required Feature Switches (Must Be Checked)

Path: Xiaozhi Assistant → Enable Wake Word Detection in Listening Mode After enabling, device will continuously receive audio and monitor for wake word, otherwise wake function will not work.