rnagabh/gemma4-audio-encoder

Gemma 4 Audio Encoder (USM-style Conformer)

Standalone extraction of the audio encoder from Google's Gemma 4 multimodal model family. This is a 304.8M parameter USM-style Conformer that converts audio waveforms (via 128-bin mel spectrogram) into embeddings.

License: Apache 2.0 (inherited from Gemma 4 — no restrictions)

Architecture

Property	Value
Total parameters	304.8M
Architecture	USM-style Conformer (Macaron-net)
Hidden dimension	1024 (pure audio representation)
Output dimension	1536 (text-projected via output_proj)
Conformer layers	12
Attention heads	8 (128 dim per head)
FFW intermediate	4096 (4× expansion)
Depthwise conv kernel	5
Subsampling conv channels	[128, 32]
Input	128-bin mel spectrogram @ 16kHz
Conformer activation	SiLU
Subsampling activation	ReLU
Conformer normalization	RMSNorm (eps=1e-6)
Subsampling normalization	LayerNorm
Residual weight	0.5 (Macaron half-step)
Attention type	Chunked causal (chunk_size=12, left_context=13, right_context=0)
Clipped linears	Yes (quantization-ready input_min/max, output_min/max per layer)
Temporal downsampling	4× (two stride-2 Conv2d layers)

Conformer Block Structure

Each of the 12 conformer blocks follows the Macaron-net pattern:

Input
  → FFW1: pre_layer_norm → Linear(1024→4096) → SiLU → Linear(4096→1024) → post_layer_norm
  → + 0.5 × residual
  → Self-Attention: norm_pre_attn → Q/K/V proj (1024→1024) → relative position → post proj → norm_post_attn
  → + residual
  → LightConv1d: pre_layer_norm → Linear(1024→2048, gated) → DepthwiseConv1d(k=5) → conv_norm → Linear(1024→1024)
  → + residual
  → FFW2: pre_layer_norm → Linear(1024→4096) → SiLU → Linear(4096→1024) → post_layer_norm
  → + 0.5 × residual
  → norm_out

Input/Output Shapes

• Input: (batch, time_frames, 128) — 128-bin mel features, time-first
• Output: (batch, time_frames/4, 1536) — 4× temporal downsampling, projected to 1536
• For 4 seconds of 16kHz audio: input ~(1, 399, 128) → output ~(1, 100, 1536)

Usage

import torch
import numpy as np
from transformers import Gemma4AudioModel, Gemma4AudioFeatureExtractor

# Load audio encoder directly from this repo
audio_tower = Gemma4AudioModel.from_pretrained(
    "rnagabh/gemma4-audio-encoder",
    torch_dtype=torch.bfloat16,
)
audio_tower.to("cuda")
audio_tower.eval()

# Load feature extractor (saved in this repo)
feature_extractor = Gemma4AudioFeatureExtractor.from_pretrained("rnagabh/gemma4-audio-encoder")

# Extract features from audio
waveform = np.random.randn(64000).astype(np.float32)  # 4s @ 16kHz
inputs = feature_extractor([waveform], sampling_rate=16000, return_tensors="pt")

with torch.no_grad():
    mel = inputs["input_features"].to(dtype=torch.bfloat16, device="cuda")

    # === Option 1: Text-projected embeddings (1536-dim) ===
    # Use this if feeding into an LLM or need the full model output.
    output = audio_tower(mel)
    text_projected = output.last_hidden_state  # (1, 100, 1536)

# === Option 2: Pure audio embeddings (1024-dim) ===
# Captures the conformer output BEFORE the text projection layer.
# Recommended for downstream audio tasks (classification, verification, etc.)
# Note: this registers a hook and runs a separate forward pass.
pre_proj_features = {}
def hook_fn(module, input, output):
    pre_proj_features["hidden"] = input[0]

handle = audio_tower.output_proj.register_forward_hook(hook_fn)
with torch.no_grad():
    _ = audio_tower(mel)
handle.remove()
audio_embeddings = pre_proj_features["hidden"]  # (1, 100, 1024)

Which to use? For audio-only tasks (classification, speaker verification, deepfake detection), the 1024-dim pre-projection embeddings are better — they retain acoustic detail that the text projection discards. The 1536-dim output is designed for feeding into an LLM decoder.

Critical: The AutoModel Loading Gotcha

⚠️ AutoModel.from_pretrained("google/gemma-4-E2B-it") silently fails to load audio tower weights.

All audio tower parameters initialize as random (std ≈ 0.02). The model runs without errors, produces outputs of the correct shape, but the outputs are meaningless.

Root cause: The checkpoint stores keys with a model. prefix (e.g., model.audio_tower.layers.0...). AutoModel builds the module tree expecting keys without the prefix. The mismatch causes every key to be both UNEXPECTED and MISSING. Transformers loads with strict=False by default, so this silently initializes everything fresh.

Fix: Use AutoModelForMultimodalLM instead:

# ❌ WRONG — audio tower weights are randomly initialized
model = AutoModel.from_pretrained("google/gemma-4-E2B-it")
audio_tower = model.audio_tower  # RANDOM WEIGHTS

# ✅ CORRECT — audio tower weights load properly
model = AutoModelForMultimodalLM.from_pretrained("google/gemma-4-E2B-it")
audio_tower = model.model.audio_tower  # TRAINED WEIGHTS

How to verify:

w = audio_tower.output_proj.weight.float()
print(f"std={w.std().item():.6f}")
# ✅ Trained: std ≈ 0.031250
# ❌ Random:  std ≈ 0.019884

E2B and E4B Share Identical Audio Weights

The audio encoder weights are byte-for-byte identical between Gemma 4 E2B and E4B. This was verified empirically — all 751 parameter tensors match exactly.

Gemma 4's E2B is a MatFormer sub-model nested inside E4B. The MatFormer architecture only affects the text decoder's feed-forward dimensions. The audio tower sits outside the MatFormer nesting and is a shared module.

Implication: There is no reason to prefer E4B over E2B for audio encoder extraction. E2B is a smaller download (~10GB vs ~16GB).

Files in This Repo

File	Description	Size
config.json	Audio tower config (Gemma4AudioConfig)	<1 KB
model.safetensors	Audio tower weights (304.8M params, BF16)	609.7 MB
preprocessor_config.json	Mel spectrogram feature extractor config	<1 KB
embed_audio.safetensors	Audio→text embedding projection (1536→1536)	4.7 MB

Limitations

• End-to-end trained for LLM decoding: The encoder was trained to produce features for Gemma 4's text decoder, not as a general-purpose audio encoder. For standalone feature extraction, the 1024-dim pre-projection output (before output_proj) may be more useful than the 1536-dim post-projection output.
• Causal chunked attention: The encoder uses right_context=0, meaning it cannot look ahead. This limits its use in offline/non-streaming settings compared to bidirectional encoders.
• Multi-layer fusion doesn't help: Unlike wav2vec2/W2v-BERT where combining multiple hidden layers improves downstream performance, this encoder's Macaron half-step residuals and causal attention mean only the final layer output is useful.
• Subsampling frontend uses ReLU + LayerNorm (not SiLU + GroupNorm as in some USM descriptions).
• Not a speaker encoder: While embeddings show some speaker separation (cosine similarity gap of ~0.03), this model was not trained for speaker verification. Dedicated speaker models will significantly outperform it on speaker tasks.

Benchmark Results (frozen 1024-dim embeddings, linear probe)

Speech Commands Classification (35 classes)

Metric	Value
Linear probe accuracy	72.0%
Random baseline	2.9%
Improvement over chance	25×
Dataset	Google Speech Commands v0.02 (validation)
Probe	Logistic regression on L2-normalized mean-pooled embeddings

The encoder captures rich phonetic and semantic content — strong on acoustically distinct words (seven: 0.93 F1, house/stop/eight: 0.89 F1) and weaker on similar-sounding pairs (three/tree).

Speaker Similarity (LibriSpeech test-clean)

Metric	Value
Same-speaker cosine similarity	0.656 ± 0.147
Different-speaker cosine similarity	0.622 ± 0.132
Separation gap	0.034

Modest speaker separation — expected since this is an ASR-oriented encoder, not a speaker verification model.

t-SNE Speaker Clustering

Embeddings show partial speaker clustering — the encoder captures speaker characteristics as a byproduct of ASR training, but is not optimized for speaker discrimination.

Extraction Details

• Extracted from google/gemma-4-E2B-it using AutoModelForMultimodalLM
• Weights saved in BF16 as safetensors
• Forward pass verified: extracted model produces outputs with 0.0 max absolute difference from the original
• All architecture specs independently verified against the live model

References

• Gemma 4 on HuggingFace
• Gemma 4 Blog Post
• Google USM Paper — "Google USM: Scaling Automatic Speech Recognition Beyond 100 Languages"