Add model card: usage example and measured performance (M4 Max)
Thanks for publishing these whisper-tiny exports — the fixed-window two-signature layout is easy to build on.
This PR fills in the empty README with a usage example and measured performance, in the same format as the moonshine-tiny card. Both CPU variants transcribe a real 4.8 s test clip fully correctly on an Apple M4 Max — i8: encode 35.7 ms, 12.0 ms per decoded token, RTF 0.009 against the fixed 30 s window (ai-edge-litert 2.1.6, XNNPACK, 4 threads, median of 10 runs). The 19 SoC-specific variants are listed but not measured here.
Happy to adjust the format or trim anything that doesn't fit how you want these cards to read.
Added a Galaxy S26 section since opening this: encoder medians for the NPU (on-device JIT and the SM8850-precompiled file) and the GPU, same conditions discipline as the M4 table. Notables: the GPU is the fastest unit on this encoder (15.0 ms against 33.7 ms on the NPU), the i8 file does not compile on the GPU delegate, and the precompiled file loads in ~150 ms against a one-time 25 s JIT compile.