NOTICE

EmbeddingGemma-2-Transcribe-CoreAI — NOTICE

This repository redistributes, in Apple Core AI `.aimodel` / `.aimodelc` form, two graphs licensed under the Apache
License, Version 2.0 (see LICENSE):

- eg2_audio_fp16_T3000: the audio encoder of EmbeddingGemma 2 (google/embeddinggemma-2, revision
  914f7f89142e33e77833254d9c9b90c3cef7303b, Google DeepMind; the Gemma 4 audio tower and embed_audio), licensed under
  the Apache License, Version 2.0. Deployments of EmbeddingGemma 2 must adhere to the Gemma Prohibited Use Policy
  (https://ai.google.dev/gemma/prohibited_use_policy). config.json is the model's configuration at that revision,
  unchanged.
- eg2_reader_v2_fp16_T750: reader v2 of EmbeddingGemma 2 Transcribe (mlboydaisuke/EmbeddingGemma-2-Transcribe-Reader),
  trained by mlboydaisuke (john-rocky), 2026-10, with CTC on LibriSpeech train-clean-100, train-clean-360 and
  train-other-500 (V. Panayotov, G. Chen, D. Povey, S. Khudanpur, "Librispeech: an ASR corpus based on public domain
  audio books", ICASSP 2015, https://www.openslr.org/12, CC BY 4.0), together with a word table of EmbeddingGemma 2
  text embeddings of the words of the LibriSpeech lexicon (https://www.openslr.org/11, public domain); vocab.txt lists
  those words.

Conversion: mlboydaisuke (john-rocky), 2026-10. Code, gates and records:
https://github.com/john-rocky/coreai-model-zoo (conversion/eg2_transcribe, models/embeddinggemma-2-transcribe).
Modifications to the encoder as exported: a fixed 30 s window (3,000 log-mel frames, 750 tokens); float16 weights and
arithmetic; every RMSNorm divides its row by the row's maximum absolute value before squaring (exact in real
arithmetic); the attention written in its banded form with the query and key scales folded into the projections;
the ±1e10 clamps that are no-ops at inference removed; the graphs' MLIR debug locations removed. The reader graph holds
the reader and the word table in float16 and returns the arg-max word, its logit and the blank logit per token.
The design of the banded attention, the folded scales and the max-abs RMSNorm follows the published description of
EverPaper's Core AI port (EverPaper/embeddinggemma-2-coreai v1.2, Apache-2.0); no code was copied.

Keep this NOTICE together with the LICENSE file when redistributing.