Updated 13 min 10 sec ago
October 3, 2026 19:52
RIYADH — NVIDIA announced it has trained its Nemotron 3.5 ASR multilingual automatic speech‑recognition model using the Saudi Audio Dataset for Arabic, a resource created by the Saudi Data and AI Authority together with the Saudi Broadcasting Authority.
According to a technical study posted on NVIDIA’s Developer Blog, incorporating the dataset almost halved the comprehension errors for Saudi dialects.
The model can transcribe speech in roughly 40 languages and dialects, including Arabic. Initially, it misidentified about 55 words out of every 100 spoken in Saudi dialects.
After training on 133.7 hours of Najdi and Hijazi audio drawn from the dataset, the word‑error rate fell to around 30 %.
Across the full multi‑dialect collection, the error rate dropped from 58.8 % to 35.6 %, while character‑level errors declined from 31.6 % to 12.2 %. Performance also improved for Modern Standard Arabic and English, and the training required only four and a half hours on two GPUs.
Optimized for real‑time interactive systems, the refined model achieves latency as low as 80 milliseconds, making it suitable for intelligent voice assistants, conversational agents, live‑broadcast subtitling, and transcription of call‑center recordings, media archives, and meetings.
NVIDIA has released the workflows and tools needed for developers worldwide to duplicate this approach for other languages.
The Saudi Data and AI Authority previously made the Saudi Audio Dataset for Arabic available on Kaggle to aid international researchers and developers.
The corpus comprises roughly 667 hours of transcribed audio, including more than 600 hours supplied by the Saudi Broadcasting Authority from 57 television programs and series that cover over 10 Saudi dialects, with an additional 20 hours set aside for validation.
Containing more than 125,000 categorized clips, the dataset allowed NVIDIA to focus on specific Najdi and Hijazi speech patterns.
The Saudi Audio Dataset for Arabic initiative supports academic and technical work to build advanced audio models for speech recognition, text‑to‑speech synthesis, speaker diarization, and demographic classification, while also enriching digital Arabic content.
NVIDIA’s use of the dataset underscores how high‑quality national data can help global AI systems better grasp local dialects and improve technologies tailored for Arabic speakers.
