Microsoft VibeVoice-ASR tested on Mac for a one-hour podcast

微软 1 月开源的 VibeVoice-ASR 语音识别模型（https://t.co/CsXftlEAul），Simon Willison 在 Mac 上测试后给出了一份具体的实测报告。

Simon Willison ran 4-bit VibeVoice-ASR on an M5 Max MacBook Pro and transcribed a one-hour podcast in 8m45s. The 9B MIT-licensed model supports 60-minute audio, 50+ languages, and structured speaker output. Memory is the constraint: prefill peaked at 61.5GB, making 32GB laptops impractical.

Why it matters: HKR-H/K/R all pass: Simon Willison’s local test gives speed, parameter size, and memory peak that practitioners can act on. It is a single benchmark, not a fresh model launch, so it stays at the featured threshold.