On 30 September 2026 NVIDIA published a technical post on fine-tuning its Nemotron 3.5 ASR speech recognition model for Saudi dialects. The training data was SADA, the Saudi Audio Dataset for Arabic, released by the Saudi Data and AI Authority (SDAIA) with the Saudi Broadcasting Authority. Saudi media picked the story up on 3 October.
The result is a real improvement, and it also shows how far dialect speech-to-text still is from a clean transcript. That second part matters for anyone in the Gulf planning to turn call recordings or meetings into searchable text.
The numbers, from NVIDIA's own post
- After light cleanup, the team kept 103,559 of 125,490 utterances, or 133.7 hours of Najdi and Hijazi speech.
- Word error rate (WER) on the Najdi and Hijazi test split fell from 55.05% to 29.96%. On the full SADA test set it fell from 58.84% to 35.61%.
- English did not get worse: WER on the FLEURS English set moved from 11.04% to 10.42%, and on FLEURS Arabic from 12.67% to 11.41%.
- Training took about 4.5 hours on two GPUs over 12,000 steps.
Look at the gap in that list. The same base model scored 12.67% on FLEURS Arabic, which is read, standard speech, and 55.05% on Najdi and Hijazi. The dialect is the hard part, not "Arabic". When a vendor says its product supports Arabic, ask which Arabic it was tested on.
What a 30% word error rate means in practice
WER counts words that are wrong, missing or extra. At about 30%, a ten-word sentence will often have around three of them. A reader can usually still follow the topic. The trouble is the words a support call or meeting depends on: names, order numbers, amounts, dates. Those are exactly the words you cannot guess back from context.
NVIDIA also measured a trade-off. A wider attention context cut WER by 1.31 points with no retraining, at the cost of roughly 800 ms more delay. The post says that suits batch transcription of call archives, media and meeting recordings, and may not suit live captioning.
Five rules for using dialect transcripts
- Test on your own audio. Have a person transcribe a sample of your real calls and compare. NVIDIA's own advice is to evaluate each capability on an independent set, not to trust a headline number.
- Prefer batch over live when accuracy matters. If nobody reads the transcript in real time, take the slower, more accurate setting.
- Keep the link back to the audio. Every transcript segment should carry its timestamp, so anyone who doubts a word can listen to that second.
- Treat transcripts as searchable drafts, not records. A person confirms names, numbers and decisions before they go into a ticket, minutes or a contract file.
- Make search answers cite the source. If an assistant answers from transcripts, the answer has to point to the exact file and passage, and say so when nothing supports it. We explain why in cite or refuse.
Where FanMind fits
FanMind is Fanpino's document assistant. It answers staff questions from an organisation's own files and cites the exact document and page for each answer. It does not transcribe audio, and we are not claiming dialect support here. What it does fit is the step after transcription: once call summaries or meeting minutes exist as Word or PDF files, they can sit next to policies and procedures, and every answer drawn from them links to the file and page a person can check. That is the same approach we use for technical document archives. See FanMind.
FAQ
What did NVIDIA announce on 30 September 2026?
A fine-tuning recipe for Nemotron 3.5 ASR on Saudi Najdi and Hijazi speech from SDAIA's SADA dataset, cutting word error rate on that test split from 55.05% to 29.96%.
Is Saudi dialect speech recognition solved?
No. About 30% WER on the target dialects, and 35.61% on the full SADA test set, is much better than before but still far from a transcript you can file without checking.
Should call centres use live or batch transcription?
For records and search, batch. NVIDIA reports that a wider context improves accuracy at about 800 ms of extra delay, which suits archives and meeting recordings more than live captions.
Does FanMind transcribe Arabic audio?
No. FanMind works on documents. Transcripts produced by a speech recognition system can be added as files, and answers then cite the transcript file and page.
