---
format: "aidr-story-markdown/v1"
id: "a7a75b8f97ef77627f5dfdfc54faa278ceb67ce613917ef6c07324fc7b5555f1"
canonical_url: "https://aidr.today/a7a75b8f?lang=en"
title: "How Far Can the Gemma 4 Speech Encoder Go Without Changing Its Weights? Saudi and Gulf Arabic Speech Recognition with a Frozen and an Adapter-Augmented Encoder"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-10-22T04:00:00.000Z"
category: "Research"
topics: ["google","inference"]
source_urls: ["https://arxiv.org/abs/2610.28673"]
summary: "A study evaluates the 305M-parameter audio encoder used by Gemma 4 and EmbeddingGemma 2 for Saudi and Gulf Arabic speech recognition without modifying its weights. A frozen system with a 22.6M-parameter CTC reader runs 3.4 times faster on CPU but trails in word error rate, while an adapted system with 3.1M low-rank adapter parameters and a 41.5M-parameter reader performs comparably or better on several metrics against larger models."
---

# How Far Can the Gemma 4 Speech Encoder Go Without Changing Its Weights? Saudi and Gulf Arabic Speech Recognition with a Frozen and an Adapter\-Augmented Encoder

> [Open the canonical story](<https://aidr.today/a7a75b8f?lang=en>)

**Published:** 2026-10-22T04:00:00.000Z
**Category:** Research
**Topics:** google, inference

## Summary

A study evaluates the 305M\-parameter audio encoder used by Gemma 4 and EmbeddingGemma 2 for Saudi and Gulf Arabic speech recognition without modifying its weights\. A frozen system with a 22\.6M\-parameter CTC reader runs 3\.4 times faster on CPU but trails in word error rate, while an adapted system with 3\.1M low\-rank adapter parameters and a 41\.5M\-parameter reader performs comparably or better on several metrics against larger models\.

## Sources

- [Story source](<https://arxiv.org/abs/2610.28673>)