NOTICE

V1 speech recognition on Core ML
Copyright Harshit B.

The repository as a whole is licensed under the Apache License, Version 2.0 (LICENSE).

1. Core ML package (v1.mlpackage), derived from google/embeddinggemma-2

   It contains layers 0-11 of the audio tower of EmbeddingGemma 2 by Google DeepMind,
   https://huggingface.co/google/embeddinggemma-2 (revision 914f7f89142e33e77833254d9c9b90c3cef7303b),
   licensed under the Apache License, Version 2.0
   (https://ai.google.dev/gemma/docs/gemma_4_license).

   Modifications made by Harshit B.:
   - the audio tower truncated after layer 11 and converted to Core ML ML programs,
     one function per fixed input window (5, 10, 20, 35 s) sharing one set of weights;
   - blocked local attention rewritten as dense banded attention (identical in fp32);
   - RMSNorm inputs scaled by powers of two so fp16 norms on the CPU / GPU do not
     overflow, and constant padding fill (identical in fp32);
   - log-softmax rewritten in a numerically stable form; static shapes;
   - a per-frame int8 quantize/dequantize step, standardization and a CTC head appended;
   - weights stored in float16.

   The upstream model card states that deployments must adhere to the Gemma Prohibited
   Use Policy: https://ai.google.dev/gemma/prohibited_use_policy

   config.json embeds the upstream preprocessor_config.json unchanged.

2. CTC head (inside v1.mlpackage) and tokenizer (spm_unigram256.model)

   Trained by Harshit B. on the LibriSpeech ASR corpus, OpenSLR SLR12,
   https://www.openslr.org/12/ (license stated there: CC BY 4.0).
   V. Panayotov, G. Chen, D. Povey and S. Khudanpur, "LibriSpeech: an ASR corpus based
   on public domain audio books", ICASSP 2015.

3. Language model (sp256_6gram.bin)

   Built with KenLM (https://github.com/kpu/kenlm; no KenLM code is included) from
   librispeech-lm-norm.txt.gz of "LibriSpeech language models, vocabulary and G2P models",
   OpenSLR SLR11, https://www.openslr.org/11/ (license stated there: Public domain).

4. Decoder (not included; used by the usage example)

   ctc-beam-decoder 0.2.0, https://pypi.org/project/ctc-beam-decoder/,
   Apache License, Version 2.0.