Nvidia reports faster, more accurate Saudi Arabic transcription in decoder test

Nvidia reports that changing the decoding configuration of a fine-tuned Nemotron speech recognition model improved Saudi Arabic transcription accuracy while reducing runtime in its test. Using NeMo's batched MALSD beam-search strategy, beam 4 lowered the word error rate from 29.96% to 28.81% at 0.59× the greedy runtime. The workflow specializes the model for Najdi and Hijazi speech and mixes previously learned English and Arabic data into training to limit forgetting. A separate change that lets the model consider more future audio reduced word error rate by 1.31 absolute points without retraining, but added approximately 800 ms of latency. Nvidia limits its decoder conclusion to MALSD and says the workflow does not establish results for every Arabic dialect or deployment environment.