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

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.