Architecture

flowchart LR
    NVDA[NVDA or SAPI speech] --> Speakers[Windows default render device]
    Speakers --> Loopback[WASAPI loopback capture]
    Loopback --> Wav[16 kHz 16-bit mono PCM WAV]
    Wav --> Whisper[Local Whisper.net ggml model]
    Whisper --> Text[Transcribed text]
    Text --> Score[Token Jaccard + ordered Levenshtein assertion]
    Expected[Expected text] --> Score

RoundTripService owns orchestration and depends only on IAudioCapture, ISpeechSynth, and ITranscriber. Unit tests mock those seams. Windows-specific adapters live at the edge:

• WasapiCapture: NAudio WASAPI loopback capture and WAV conversion.
• SapiSpeech: System.Speech/SAPI synthesis.
• WhisperTranscriber: lazy Whisper.net model load and local transcription.

Tool surface

graph LR
    C[Claude / MCP client] -->|stdio| S[AudioLoopMcp.Server]
    S --> T1[speak]
    S --> T2[record_loopback]
    S --> T3[transcribe]
    S --> T4[list_audio_devices]
    S --> T5[round_trip_test]
    S --> T6[listen_and_transcribe]
    T1 --> SAPI[System.Speech SAPI]
    T2 --> WAS[NAudio WASAPI loopback]
    T3 --> WH[Whisper.net local model]
    T5 --> SAPI
    T5 --> WAS
    T5 --> WH
    T6 --> WAS
    T6 --> WH