Experiment 9-1 real traditional-voice validation

• Run ID: exp9-1-20260729T153334474Z
• Complete: true
• Source: microphone_input.wav (5.056 s, saved browser microphone/WebSocket capture)
• VAD: Silero ONNX, 500 ms silence, non-forced endpoint = true
• ASR: local-openai-whisper / whisper-tiny, 2.710 s
• Transcript: 是男朋友叫这样的
• LLM: ark / doubao-seed-1-6-flash-250615, TTFT 0.749 s, total 0.792 s
• Response: 那是男朋友叫这样的呀
• TTS: fish / s1, first byte 2.819 s, total 3.279 s
• Post-endpoint time to first audio byte: 6.279 s

Strict gates

• schema_and_scope: true
• real_websocket_microphone_media: true
• real_silero_vad_endpoint: true
• real_asr: true
• real_streaming_llm: true
• real_tts_media: true
• measured_stage_latencies: true
• provenance_complete: true
• no_mock_probe_or_fallback: true

Passing proves one saved real microphone turn completed Silero VAD -> real ASR -> real LLM -> real TTS. It does not benchmark concurrency or production load.