zipformer2-66m-streaming-he - Hebrew speech recognition model used by /tools/transcribe/

Author: Yakov Kolani (original: https://huggingface.co/thewh1teagle/zipformer2-66m-streaming-he),
mirrored with extra exports by ivrit.ai: https://huggingface.co/ivrit-ai/zipformer2-66m-streaming-he
Licence: CC-BY-4.0 (LICENSE-CC-BY-4.0.txt). Base model: k2-fsa/icefall streaming Zipformer (Apache-2.0).
Training audio: ivrit.ai (https://www.ivrit.ai/en/the-license/).

Files (fetched 2026-10-19 from the ivrit-ai repo, sherpa-onnx/ and onnx/ folders, unchanged):
  sherpa-onnx/encoder.int8.onnx  sha256 6c6febcb74a09435315a1d872a6539f6e7fc2c333533f14c74edc0e85c49ffcd  70,478,964 bytes
    -> split -b 25165824 -d -a 1 into encoder.int8.onnx.part0..2 (Cloudflare Pages caps a file at 25 MiB);
       the browser joins the parts and caches them (Cache API "nayeret-models-v1").
  sherpa-onnx/decoder.int8.onnx  sha256 feb03e7c8aa761593a9284a6fa1e91259be3702de75f630062256c1a59c5a906
  sherpa-onnx/joiner.int8.onnx   sha256 1d2c6d8e6dabc5eb52de52e95718e9bdfbffbfe7ead3808fb5ffd74e83bd9cae
  sherpa-onnx/tokens.txt         sha256 e4ebb5e72a6aeef47437a55fa2308ebbedb7c51c557de7a89bf965c23ea9af4a
  onnx/config.json               sha256 6ad75405069d306cf96f1642e8093f7642773f40082c39a0ce7a3d270b1db920 (re-serialised without whitespace)

Inference code is ours (src/js/workers/asr-core.mjs: Kaldi-style fbank + streaming encoder + greedy transducer search),
checked against sherpa-onnx 1.13.8 on the same clips (research/transcription-spike.md).