How SherpaOnnx Powers cueflow's 2-pass ASR / 我们是如何利用sherpa-onnx实现CueFlow的ASR功能

[wanghsinche]

We want to give a great thanks to the SherpaOnnx team for creating such an amazing ASR framework! 🎉

非常感谢 SherpaOnnx 团队！

In our cueflow project (website), we're using SherpaOnnx as the core of our audio processing pipeline to handle:
✅ Dual audio sources (mic + system audio)
✅ Real-time streaming with online workers
✅ High-accuracy batch processing with offline workers

Here's how we've integrated SherpaOnnx into our architecture:

graph TD
    subgraph "Audio Input"
        Mic["Microphone Input"]
        SysAudio["System Audio"]
    end

    subgraph "Audio Processing"
        StreamMgr["Audio Stream Manager"]
        SherpaSvc["SherpaOnnx Service"]
        OnlineWorker["Online ASR Worker <br/>(Streaming)"]
        OfflineWorker["Offline ASR Worker <br/>(VAD)"]
    end

    subgraph "Final Processing"
        FinalProcessing["Transcript Fusion <br/> AI Analysis"]
    end

    %% Simplified Data Flow with Online/Offline Distinction
    Mic --> StreamMgr
    SysAudio --> StreamMgr
    StreamMgr --> SherpaSvc
    SherpaSvc -- "Audio" --> OnlineWorker
    SherpaSvc -- "Audio" --> OfflineWorker
    OnlineWorker -- "Realtime Transcripts" --> SherpaSvc
    OfflineWorker -- "Final Transcripts" --> SherpaSvc
    SherpaSvc --> FinalProcessing

    %% Styling
    style StreamMgr fill:#ccf,stroke:#333,stroke-width:2
    style SherpaSvc fill:#fcc,stroke:#333,stroke-width:2
    style OnlineWorker stroke:#333,stroke-width:1,stroke-dasharray: 5 5
    style OfflineWorker stroke:#333,stroke-width:1,stroke-dasharray: 2 2
    %% Removed style for FinalProcessing

Loading

Some key benefits we've experienced:
⚡ Low-latency performance for real-time use cases
🔍 Excellent accuracy in 2-pass streaming and offline modes
🧩 Easy to integrate with our existing Electron/Node.js stack

We'd love to hear from other SherpaOnnx users - happy to answer any questions about our implementation in cueflow github issues!