Analog Voice Activity Detector (VAD) is a promising candidate for a power and cost-efficient solution for AIoT voice assistants. Regrettably, the PVT variation from the analog circuits and data misalignment from sensors limit the VAD accuracy with conventional backpropagation model-based training (BPMBT). This brief presents a forward-forward closed box trainer (FFBBT) for analog VADs. It trains the analog circuit without knowing the circuit model or finding its gradient. Thus, it is insensitive to PVT variation and offset, achieving a measured VAD accuracy improvement of ∼3% and an accuracy variation reduction of 5.6×. Moreover, a Tensor-Compressed DerivativeFree Optimizer (TCDFO) is also proposed to reduce the required memory for FFBBT by 1600×. The FFBBT with TCDFO is synthesized in 28 nm CMOS with a power of 512 nW and an area of 0.003 mm².