AI article

Whisper on the RK3588 NPU: the encoder offload works — everything around it is the real lesson

Community description: Status: encoder offload merged into a working whisper.cpp path on an RK3588 board, transcription...

Dev.to | Nov 9, 2026 | Anguo

Automated excerpt

RKNN encoder, single run: 703.0 ms (of which rknn_run 676.8 ms) End-to-end transcription: 13.76 s → 11.15 s, output byte-identical with the offload on and off The glue code is 143 lines. The Model Zoo build changed CHUNK_LENGTH to 20 s; its output frame count doesn't match whisper.cpp's native 30 s / 1500 frames. CPU baseline first → the encoder was the single biggest item (9.57 s median encode on an 11 s clip, loaded board).

Selected automatically from source text; not independently written or fact-checked. Read the original for full context.

Read the original article

Read next

AI briefing: recent picks

More stories to explore