Nvidia improves AI speech recognition for Saudi dialects

Nvidia has significantly improved its Nemotron 3.5 ASR speech recognition model by utilizing the ‘Sada’ audio dataset, which was released by the Saudi Data and AI Authority (SDAIA) in collaboration with the Saudi Broadcasting Authority.

By training the model on approximately 133.7 hours of Najdi and Hejazi dialect recordings, Nvidia reduced error rates for these specific dialects from roughly 55% to approximately 30%. The overall error rate for the entire Sada dataset also dropped from 58.8% to 35.6%, while character-level errors fell from 31.6% to 12.2%. The training process was highly efficient, taking only about 4.5 hours using two GPUs.

The Nemotron 3.5 ASR model supports real-time transcription for approximately 40 languages and dialects. This improvement makes the technology better suited for interactive applications such as smart voice assistants, conversational agents, live broadcast transcription, and call center archiving. The training did not negatively impact the model's performance in English or Modern Standard Arabic, showing slight improvements in those areas as well.