Whistle at a glance
INPUT
16 kHz mono audio, up to 30 seconds in one pass
OUTPUT
the transcript, the language, per-word times, and the encoder embedding
MODEL
16.9 MB at 2 to 4 bit: log-mel front end, conv stem, encoder, laddered decoder
SHARED
Needle's .cact container, Cactus Quants, SIMD kernels and KV cache
LANGUAGES
English, German, French, Spanish, Italian, Dutch and Polish
SPEED
11.1 ms to the first token, 1,319 tokens/s, on an Apple M4 Pro CPU
LADDER
every decoder depth from 2 layers up is deployable, picked with --audio-depth