ArcForge Automatic Speech Recognition (ASR) SDK

High-Performance Embedded ASR Inference Framework based on Modern CMake

English | 简体中文

---

📖 Introduction

ArcForge is a C++ ASR inference framework designed specifically for edge computing. It integrates the sherpa-zipformer model, implementing a complete pipeline from low-level NPU binding to high-level Client/Server calls. The standout feature of this project is "Architectural Transparency" — we fully disclose and deconstruct the complete industrial-grade Modern CMake build logic.

🏗️ Architecture

We provide not just the source code, but also a fully visualized build system.

🔍 Deep Dive: ArcForge Build System Explained (BUILD_SYSTEM.md) (Click the link above to view the Build Lifecycle, Core DSL Logic, and Detailed Dependency Topology)

Core Module Overview:

graph LR
    %% ==========================================
    %% High Contrast Styles
    %% Strategy: Dark background + White text + Bright borders
    %% ==========================================
    classDef appNode fill:#0d47a1,stroke:#90caf9,stroke-width:2px,color:#fff;
    classDef libNode fill:#e65100,stroke:#ffcc80,stroke-width:2px,color:#fff;
    classDef thirdNode fill:#333333,stroke:#ffffff,stroke-width:2px,color:#fff,stroke-dasharray: 5 5;

    subgraph Apps [Application Layer]
        direction TB
        Client(ASR Client):::appNode
        Server(ASR Server):::appNode
    end

    subgraph Libs [Core Library Layer]
        direction LR
        Engine[ASREngine]:::libNode --> Net[Network]:::libNode
        Net --> Utils[Utils]:::libNode
    end

    subgraph 3rd [Dependency Layer]
        direction LR
        Engine -.-> Sherpa[SherpaOnnx]:::thirdNode
        Sherpa -.-> RKNN[RKNNRT]:::thirdNode
    end

    Client --> Engine
    Server --> Engine

Loading

---

🚀 Quick Start

This project relies heavily on .env for environment isolation. Please strictly follow the steps below:

1. Environment Configuration (Required)

# 1. Copy the configuration template
cp .env.example .env

# 2. Edit the configuration file
#    ARC_RV1126BP_SDK_ROOT: Path to your cross-compilation toolchain root (Required)
#    ARC_DEP_CACHE_DIR:     Path to third-party library cache (Recommended for faster builds)
vim .env

2. Build

Use the build.sh script to drive CMake Presets:

# [Recommended] Clean Build (RV1126 Release)
./build.sh cb rv1126bp

# [Dev] Incremental Build (Compile only modified parts)
./build.sh build rv1126bp

# [Debug] Native Verification (x86 Debug)
./build.sh cb native debug

---

⌨️ Command Cheatsheet

Scenario	Command	Description
Production	./build.sh cb rv1126bp	Clean & Build Release version (-O3)
Debugging	./build.sh cb rv1126bp debug	Build Debug version (with symbols, -Og)
Cleanup	./build.sh cleanall	Remove build/ and install/ directories
Git Clean	git clean -fdx -e .env	Caution: Remove all untracked files (keeps config)

---

📂 Output Directory

After building, artifacts are located in the install/ directory. We automatically handle renaming and RPATH settings, so they are ready for deployment:

install/rv1126bp/release/
├── bin/
│   ├── ArcForge_ASR_Client      # Client test application
│   └── ArcForge_ASR_Server      # Server main application
└── lib/
    ├── libarcforge_asrengine.so # Core engine library
    ├── libarcforge_network.so   # Network library
    ├── libarcforge_utils.so     # Basic utilities library
    └── libsherpa-onnx*.so       # Dependency library

🧩 Supporting New Platforms

Currently, only RV1126BP and RK3588S platforms are supported. To support a new platform, there are two suggested methods:

1. Add it yourself

Please read this tutorial carefully:

🔍 Deep Dive: How to Add Support for a New SoC in ArcForge(ADD_NEW_SOC.md) (Click the link above to view detailed steps guiding you through supporting a new SoC)

2. Submit an issue to me