benchmarks
latest samples
smoothing
whistle-hu teaches Cactus Whistle (16.9 MB on-device ASR) Hungarian, for the m5whistle StickS3 port: LoRA on the attention + conv-module projections, new Hungarian BPE pieces (embedding rows), decoder cut to the depth-2 rung the device runs (layers 0 and 7), engrams off, and fake quantisation at the published archive's widths (2-bit kernels, 4-bit embedding / mHC).
Val WER / CER: greedy decoding on 300 held-out Common Voice + VoxPopuli clips, Whisper basic normalizer. Test: 1000 clips each of Common Voice test, FLEURS test, VoxPopuli test. Cost: GPU board power integrated from nvidia-smi × €/kWh (GPUs only).
stormserver:/srv/claude-offload/jobs/whistle-hu