Freezing a 305-million-parameter speech encoder and using it to transcribe Saudi and Gulf Arabic is the bet at the heart of a study that tests the encoder shared by Gemma 4 and EmbeddingGemma 2. Unlike typical announcements, the backbone is not retrained. A frozen system trains only a 22.6-million-parameter CTC reader over a mix of layers; an adapted variant adds 3.1 million low-rank adapter parameters, a 41.5-million-parameter hybrid CTC/attention reader and joint decoding with a sub-word language model. The point is not a single word error rate, but the ratio between adaptation cost and performance on dialect varieties that commercial services often cover poorly.

A rerun against Cohere Transcribe Arabic (2.07 billion parameters) on identical audio shows a split picture. The frozen system trails in WER, but on five of six human-labelled conditions the character error rate gap is statistically indistinguishable; in return it runs 3.4 times faster on CPU with greedy decoding. The adapted version holds WER on a public Saudi customer-service call (19.20% versus 19.27%) and on clean Emirati speech, is significantly better on telephone-coded Emirati speech, has significantly lower CER across all six conditions of these sets and beats a 0.6-billion-parameter recogniser fine-tuned on SADA in WER on all six. It still trails Cohere on long-form Saudi speech.

On a blind, human-corrected public call-centre sample, the system ranks fourth out of fifteen, behind three commercial APIs (only one by a statistically significant margin) and ahead of the other self-hosted systems by point estimate. The optimised joint decoder runs at 139 times real time on one GPU.

The structural reading is this: for teams managing voice data in regulated settings, a frozen encoder reduces the surface to retrain and the number of parameters to move. You do not need to bring a billion-parameter model in house to get acceptable transcription on specific dialects: a light reader and, when needed, low-rank adapters are enough. That changes the TCO calculation: adaptation does not touch the backbone weights, so the risk of regression on other varieties is contained and updates can happen component by component. CPU speed opens room for edge deployments or infrastructure without dedicated GPUs, even though the more accurate joint decoder needs a GPU to sustain 139 times real time.

The picture is not a full win. WER on Saudi long-form audio still trails the commercial API, and telephone-coded speech is the toughest test. There is also an evaluation problem: when references are created by correcting machine drafts, the system that produced the drafts is favoured. That is a known effect for annotation work, but here it matters because it can inflate the ranking of self-hosted systems if reference transcripts are not independent.

For teams evaluating on-premise deployments, the trade-off is no longer 'small model versus giant API' but 'how much adaptation is needed to bring the model into your own infrastructure'. AI-RADAR analyses these trade-offs at /llm-onpremise. The study, with code and public-set results released, offers a concrete starting point: a route for Gulf Arabic that does not pass through the cloud.