How Far Can a Frozen Gemma 4 Speech Encoder Go? Saudi and Gulf Arabic ASR with Frozen and Adapter-Augmented Encoders
arXiv:2610.28673 [cs.CL] Β· Submitted 18 October 2026 Β· 24 pages, 15 tables, 5 figures
Author: Amr Abdelkhalik
Subjects: Computation and Language (cs.CL); Sound (cs.SD); Audio and Speech Processing (eess.AS)
Code and public-set results: https://github.com/AmrKhalil9/frozen-embeddinggemma2-arabic-asr
Overview
This study investigates how far the 305M-parameter audio encoder shared by Gemma 4 and EmbeddingGemma 2 can be pushed for Saudi and Gulf Arabic speech recognition without modifying any of its weights. The encoder is identical across both models except for its final projection layer. The author evaluates two configurations: a fully frozen encoder paired with a lightweight CTC reader, and an adapter-augmented variant that adds low-rank adapters and a hybrid reader.
System Configurations
Frozen system. Only a 22.6M-parameter CTC reader is trained, drawing on a mix of encoder layers while the encoder itself remains untouched.
Adapted system. This variant adds 3.1M low-rank adapter parameters, a 41.5M-parameter hybrid CTC/attention reader, and joint decoding with a sub-word language model.
Results
Frozen system vs. Cohere Transcribe Arabic
Compared against Cohere Transcribe Arabic (2.07B parameters), re-run on identical audio, the frozen system trails in word error rate (WER) but is statistically indistinguishable in character error rate (CER) on five of six human-labelled conditions. It also runs 3.4Γ faster on CPU with greedy decoding.
Adapted system
The adapted system achieves the following:
- Statistically indistinguishable WER on a public Saudi customer-service call (19.20% vs. 19.27%) and on clean Emirati speech.
- Significantly better on telephone-coded Emirati speech.
- Significantly lower CER on all six conditions of these sets.
- Beats a 0.6B recogniser fine-tuned on SADA in WER on all six conditions.
- Still trails Cohere in WER on long-form Saudi speech.
Blind call-centre benchmark
On a blind, human-corrected public call-centre sample, the adapted system ranks 4th of 15 systems. It trails three commercial APIs (significantly only one) and leads the other self-hostable systems by point estimate.
Efficiency
An optimised joint decoder runs the adapted system at 139Γ real time on a single GPU.
Methodological Finding
References produced by correcting machine drafts tend to favour the drafting system, a bias worth accounting for in evaluation design.
Availability
Code and public-set results are released at the repository linked above.
via ArXiv CL+LG