Skip to content

Large Model Solution Development Process and Precautions*

1. Development Process*

Phase 1: Hardware Design (Strictly Follow Official Specifications)*

  1. Solution Selection
    • Audio transmission solution (choose one of two):
      • UART solution: Strong universality, low requirements on host MCU resources. Also need to select the flow control method for audio playback.
        • ACK flow control: Controls audio playback speed via message replies, low efficiency, does not support PCM format playback
        • IO flow control: Controls audio playback speed via IO level states, supports PCM playback
      • I2S solution: High requirements on host MCU resources, requires the host MCU chip to support audio codec.
  2. Hardware Design Review
  3. Acoustic Structure Design Review
    • Please carefully read the 国芯微 产品结构设计
    • Must submit the structural design to 国芯微 for review to avoid structural design issues.

Phase 2: Acoustic Debugging and Testing*

  1. Acoustic Testing
    • Carefully read the 1-声学测试用户指南 in the development kit:
      • Microphone acoustic testing: Test noise floor, sensitivity, signal-to-noise ratio, and sealing
      • Speaker acoustic testing: Test THD and AEC performance
  2. Wake-up rate testing
    • Carefully read the 3-唤醒率测试用户指南 in the development kit

Phase 3: Protocol Integration*

  1. Obtain the demo development board: Please contact our sales team.
  2. Obtain the development kit
  3. WIFI\CAT1 side development: Please carefully read the 串口协议移植指南
    • Refer to the lightning project in the gitlab repository to develop the UART driver:
      • SDK link: http://gitlab.nationalchip.com/nationalchip/voice-wifi-solution/lightning
      • Reference directories:
        • Dependent memory pool code: components/utils/mempool
        • Dependent ring buffer code: components/utils/ringbuffer
        • Audio buffer: project/ln_model_public/app/audio
        • Serial port protocol implementation: project/ln_model_public/app/smartbot
      • Parse the 国芯微 offline voice protocol, see details in GX8006开发包→4-GX8006标准固件以及串口协议移植指南→国芯微离线语音协议_v*.pdf.
      • Support dynamic configuration (microphone gain/VAD sensitivity/noise reduction level), etc.
  4. OTA upgrade
    • It is recommended to use the bare-chip upgrade method first
      • See details in GX8006开发包→4-GX8006标准固件以及串口协议移植指南→国芯微离线语音协议_v*.pdf.
    • Dual-backup OTA solution

2. Common Precautions*

2.1. Common Precautions for Acoustic Testing*

  1. Solutions for environments without an anechoic chamber (<45dBA)
    • Request professional support: If an anechoic chamber is unavailable, contact 国芯微 sales personnel to arrange a compliant acoustic testing environment.
  2. High microphone noise floor:
    • PCB circuit check:
      • Unplug the microphone. If the noise floor issue persists, the problem originates from the PCB circuit design. Request hardware intervention for optimization or contact the 国芯微 project manager for assistance.
    • Environmental noise verification:
      • Seal the microphone hole with plasticine. If the noise floor decreases significantly, environmental noise interference is significant, and the test environment's sound insulation needs improvement;
      • If the above operations are ineffective, it is recommended to replace with a far-field microphone with a high signal-to-noise ratio (e.g., ≥65dB) and high sensitivity.
  3. Key requirements for speaker acoustic testing
    • Sound pressure level limit:
      • The sound pressure level generated by the speaker at the microphone position must not exceed 88dBA, otherwise, it will cause signal clipping distortion, affecting THD and AEC performance testing.
  4. Acoustic structure precautions:

2.2. Common Precautions for Protocol Integration*

2.3. How to Customize Wake-up Words*

Model Deployment Introduction

For model deployment methods for self-trained models or models provided by 国芯微, see: Kws模型部署工具说明

  • Developers train their own models: For details, please read: Bunkws自动化训练平台搭建. This document provides two methods:
    • Method 1: Use a cloud server for training custom wake-up words + short commands
    • Method 2: Set up your own server for training custom wake-up words + short commands
  • Developers can also commission 国芯微 for training through business channels.