Thanks! On the Corstone-300 FVP (Cortex-M55 + Ethos-U55-128, int16 graphs), peak working memory is about 450 KB, excluding the stack:
- NPU arena: 270 KB. The three graphs share one arena, so this is the largest graph (the encoder, sized for a 64-syllable chunk).
- CPU buffers: 182 KB. 116 KB is the chunk state and 63 KB is the synthesiser (FFT and overlap-add buffers, streamed in 256-sample blocks, so it doesn't grow with sentence length).
Flash is separate: 218 KB of NPU graphs plus 132 KB of tables and code.