docs: add provenance / EU AI Act Art. 53 note

725806c verified 8 days ago

3.77 kB

license: apache-2.0
language:
  - en
  - multilingual
pipeline_tag: automatic-speech-recognition
tags:
  - audio
  - speech-recognition
  - gguf
  - gemma
  - conformer
library_name: ggml
base_model: google/gemma-4-E4B-it

Gemma-4-E4B-it — GGUF

GGUF conversion of google/gemma-4-E4B-it for use with CrispStrobe/CrispASR.

E4B is the larger sibling of cstr/gemma4-e2b-it-GGUF: the same gemma4 architecture (byte-identical 1024-dim USM Conformer audio tower) with a larger decoder. It runs on the same CrispASR backend with no code changes.

The Gemma-4 12B model is a different architecture (gemma4_unified, 640-dim audio encoder) and is not supported by this backend.

Available variants

File	Quant	Size	Notes
`gemma4-e4b-it-f16.gguf`	F16	~14.6 GB	Full precision
`gemma4-e4b-it-q8_0.gguf`	Q8_0	~7.8 GB	Near-lossless quant
`gemma4-e4b-it-q4_k.gguf`	Q4_K	~4.1 GB	Standard quant (auto-download default)

Model details

Architecture: USM Conformer audio encoder (12L, 1024d, chunked-local attention with relative position bias, LightConv1d, ClippableLinear with QAT scalars) + Gemma4 LLM decoder (42L, 2560d, GQA 8Q/2KV, per-layer embeddings, hybrid sliding/full attention, GeGLU)
Parameters: ~~4B effective (~~8B with embeddings)
Audio: Gemma4AudioFeatureExtractor — 128-bin mel, 16 kHz, frame_length=320, hop=160, fft_length=512, semicausal padding, log(mel + mel_floor=0.001), no normalisation
Languages: 140+ (ASR + speech translation)
License: Apache 2.0
Source: google/gemma-4-E4B-it

Usage

# Auto-download the Q4_K default and transcribe:
crispasr -m gemma4-e4b --auto-download -f audio.wav

# Or point at a local file (backend auto-detected from the GGUF):
crispasr --backend gemma4-e2b -m gemma4-e4b-it-q4_k.gguf -f audio.wav

The CrispASR backend is named gemma4-e2b (it serves the whole gemma4 E2B/E4B family).

What's included vs an upstream Gemma-4 GGUF

This GGUF is built specifically for ASR with CrispASR and includes the audio path that standard text/vision Gemma-4 GGUFs omit:

12-layer audio conformer encoder.
Gemma4MultimodalEmbedder audio→LLM adapter (embed_audio.embedding_projection, pre-projection RMSNorm).
Mel filterbank + Hann window baked into the GGUF.

Converted with models/convert-gemma4-e2b-to-gguf.py.

Provenance and EU AI Act Art. 53 note

Upstream model: google/gemma-4-E4B-it — published by google.
Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.