litert-community

/

whisper-base

Like 2

Follow

LiteRT Community (FKA TFLite) 9.32k

Thanks for documenting the signatures on the card — that is exactly what made an unambiguous encoder measurement possible.

This PR adds a measured Performance section: Galaxy S26 (SM8850) rows for both files, encode signature only, conditions inline (CompiledModel 2.2.0, 50-run medians, thermal NONE, delegate confirmed from logcat). Same shape as whisper-tiny on this device: the GPU is the fastest unit on the f32 encoder (50.0 against 58.9 ms on the Hexagon), the i8 file runs NPU-only at 3.3–3.6× the f32 time, and the NPU's one-time JIT compile (~26 s) drops to a 0.24 s cached load afterwards.

Happy to reshape to the house style.