{"version":"1","article":{"id":"saudi-nemotron-sada-dialect-benchmark-2026","slug":"saudi-nemotron-sada-dialect-benchmark-2026","language":"en","url":"https://haberpaper.com/en/news/saudi-nemotron-sada-dialect-benchmark-2026","api_url":"https://haberpaper.com/api/v1/articles/saudi-nemotron-sada-dialect-benchmark-2026?lang=en","title":"NVIDIA cuts speech-recognition error on Saudi dialect test set","summary":"In its SADA experiment, NVIDIA reports that word error rate on a Najdi and Hijazi test split fell from 55.05% to 29.96%; the result is specific to the reported test data.","category":"business","region":"world","topics":[],"published_at":"2026-10-08T20:57:50+03:00","modified_at":"2026-10-08T20:57:50+03:00","publication_basis":"publication-record","event_period":"2026-09-30","event_period_basis":"event-or-reference-period","sources":[{"name":"NVIDIA Technical Blog","url":"https://developer.nvidia.com/blog/fine-tuning-nvidia-nemotron-for-saudi-arabic-dialects-with-a-path-to-other-languages/","publishedAt":"2026-09-30","checkedAt":"2026-10-08"}],"translations":[{"language":"tr","url":"https://haberpaper.com/tr/news/saudi-nemotron-sada-dialect-benchmark-2026","api_url":"https://haberpaper.com/api/v1/articles/saudi-nemotron-sada-dialect-benchmark-2026?lang=tr"},{"language":"en","url":"https://haberpaper.com/en/news/saudi-nemotron-sada-dialect-benchmark-2026","api_url":"https://haberpaper.com/api/v1/articles/saudi-nemotron-sada-dialect-benchmark-2026?lang=en"},{"language":"ar","url":"https://haberpaper.com/ar/news/saudi-nemotron-sada-dialect-benchmark-2026","api_url":"https://haberpaper.com/api/v1/articles/saudi-nemotron-sada-dialect-benchmark-2026?lang=ar"},{"language":"fr","url":"https://haberpaper.com/fr/news/saudi-nemotron-sada-dialect-benchmark-2026","api_url":"https://haberpaper.com/api/v1/articles/saudi-nemotron-sada-dialect-benchmark-2026?lang=fr"},{"language":"es","url":"https://haberpaper.com/es/news/saudi-nemotron-sada-dialect-benchmark-2026","api_url":"https://haberpaper.com/api/v1/articles/saudi-nemotron-sada-dialect-benchmark-2026?lang=es"}],"data_types":["table"],"data":{"version":1,"blocks":[{"id":"saudi-arabic-asr-test-results","type":"table","title":{"tr":"SADA konuşma tanıma test sonuçları","en":"SADA speech-recognition test results","ar":"نتائج اختبارات التعرف على الكلام في SADA","fr":"Résultats des tests de reconnaissance vocale SADA","es":"Resultados de las pruebas de reconocimiento SADA"},"description":{"tr":"NVIDIA teknik yazısındaki farklı test kümeleri için uyarlama öncesi ve sonrası kelime hata oranları.","en":"Before-and-after word error rates for separate test sets in NVIDIA’s technical post.","ar":"معدلات خطأ الكلمات قبل التكييف وبعده لمجموعات اختبار منفصلة في المقال التقني لإنفيديا.","fr":"Taux d’erreur de mots avant et après adaptation pour des tests distincts de l’article NVIDIA.","es":"Tasas de error de palabras antes y después del ajuste en conjuntos de prueba distintos del artículo de NVIDIA."},"methodology":{"tr":"WER, ikame/silme/ekleme hatalarının kelime sayısına oranıdır; doğruluk yüzdesi değildir. Her satır ayrı test kümesidir. Sonuçlar şirketin kendi teknik deneyinden gelir ve ülke çapında genellenemez.","en":"WER is substitutions, deletions and insertions relative to word count; it is not an accuracy percentage. Each row is a separate test set. Results are from the company’s own technical experiment and are not nationwide estimates.","ar":"يقيس WER الاستبدال والحذف والإضافة نسبةً إلى عدد الكلمات؛ وليس نسبة الدقة. وكل صف لمجموعة اختبار منفصلة. النتائج من تجربة الشركة التقنية ولا تمثل تقديراً على مستوى المملكة.","fr":"Le WER rapporte substitutions, suppressions et insertions au nombre de mots ; ce n’est pas un taux de précision. Chaque ligne est un test distinct. Résultats de l’expérience technique de l’entreprise, non d’une estimation nationale.","es":"WER relaciona sustituciones, eliminaciones e inserciones con el recuento de palabras; no es un porcentaje de precisión. Cada fila es una prueba independiente. Resultados del experimento técnico de la empresa, no una estimación nacional."},"sourceUrls":["https://developer.nvidia.com/blog/fine-tuning-nvidia-nemotron-for-saudi-arabic-dialects-with-a-path-to-other-languages/"],"updatedAt":"2026-10-08T19:35:56+03:00","columns":[{"key":"test_set","label":{"tr":"Test kümesi","en":"Test set","ar":"مجموعة الاختبار","fr":"Jeu de test","es":"Conjunto de prueba"},"type":"text","language":"en"},{"key":"before","label":{"tr":"Önce WER","en":"Before WER","ar":"WER قبل التكييف","fr":"WER avant","es":"WER antes"},"type":"number","unit":"percent"},{"key":"after","label":{"tr":"Sonra WER","en":"After WER","ar":"WER بعد التكييف","fr":"WER après","es":"WER después"},"type":"number","unit":"percent"}],"rows":[{"test_set":"SADA Najdi + Hijazi target split","before":55.05,"after":29.96},{"test_set":"Full SADA","before":58.84,"after":35.61},{"test_set":"FLEURS English","before":11.04,"after":10.42},{"test_set":"FLEURS Arabic","before":12.67,"after":11.41}]}]},"data_compilations":[{"id":"saudi-arabic-asr-test-results","creator":{"@type":"Organization","@id":"https://haberpaper.com/#organization","name":"Haber Paper","url":"https://haberpaper.com/"},"license":{"@type":"CreativeWork","@id":"https://haberpaper.com/en/data-terms/v1#license","name":"Haber Paper Original Data Compilation License 1.0","url":"https://haberpaper.com/en/data-terms/v1"},"rights_scope":"Haber Paper original data arrangement, labels and methodology notes only; underlying source rights and permissions remain separate.","source_license_url":null}],"data_downloads":[{"id":"saudi-arabic-asr-test-results","formats":{"json":"https://haberpaper.com/api/v1/articles/saudi-nemotron-sada-dialect-benchmark-2026/data/saudi-arabic-asr-test-results?lang=en&format=json","csv":"https://haberpaper.com/api/v1/articles/saudi-nemotron-sada-dialect-benchmark-2026/data/saudi-arabic-asr-test-results?lang=en&format=csv"}}],"content":{"paragraphs":["A 30 September NVIDIA technical blog describes adapting the Nemotron 3.5 automatic speech-recognition model to the Najdi and Hijazi dialects used in Saudi Arabia. The team used 133.7 hours of target-dialect speech from the Saudi Audio Dataset for Arabic (SADA). The blog says the curation retained 103,559 utterances from an initial 125,490, or 82.5%. This is the company’s own training and evaluation experiment, not an independent, nationwide accuracy assessment.","NVIDIA reports that word error rate (WER) on its Najdi-Hijazi target test split fell from 55.05% to 29.96%. WER across the full SADA test set declined from 58.84% to 35.61%. WER measures substitutions, deletions and insertions relative to the number of words; it is not equivalent to saying that the model became ‘25% more accurate.’ In the same experiment, FLEURS English WER changed from 11.04% to 10.42%, and Arabic WER from 12.67% to 11.41%. These figures belong to separate test sets and should not be conflated.","The training recipe mixes target-dialect speech with weighted English and Arabic examples from FLEURS and groups utterances of similar duration to manage training computation. The best reported SADA result followed 12,000 training steps and about 4.5 hours on two GPUs. Those are NVIDIA’s experimental conditions; they do not guarantee the same cost or result with different data, model versions or infrastructure.","The post stresses that results vary by evaluation set and that a gain on one test cannot be generalized automatically to other dialects or real-world settings. Curation choices, background noise, microphones, speaking rate and transcript conventions can all affect WER. NVIDIA also reports that offline decoding options lowered WER while adding latency; a wider attention context can add about 800 milliseconds. Live captioning, call-archive transcription and meeting records therefore involve different trade-offs.","Speech recognition adapted to Saudi dialects could be a technical component for customer service, media archives or voice interfaces. The source does not say that this model has been deployed by Saudi businesses or won a commercial contract. Its measurements show how the output changed on the specified test sets under the published recipe. Real deployments would still need local user testing, error analysis, privacy and consent controls, accessibility review, operating-cost assessment and appropriate human oversight.","The table reports WER percentages and keeps test sets separate. Before-and-after values are the measurements in NVIDIA’s technical post. They are experimental results and do not claim to represent every dialect or speaker in the country. The supported conclusion is that dialect-targeted fine-tuning improved the measured test performance in this experiment; performance in actual use requires separate validation."],"key_points":[],"sections":[]},"text":"In its SADA experiment, NVIDIA reports that word error rate on a Najdi and Hijazi test split fell from 55.05% to 29.96%; the result is specific to the reported test data.\n\nA 30 September NVIDIA technical blog describes adapting the Nemotron 3.5 automatic speech-recognition model to the Najdi and Hijazi dialects used in Saudi Arabia. The team used 133.7 hours of target-dialect speech from the Saudi Audio Dataset for Arabic (SADA). The blog says the curation retained 103,559 utterances from an initial 125,490, or 82.5%. This is the company’s own training and evaluation experiment, not an independent, nationwide accuracy assessment.\n\nNVIDIA reports that word error rate (WER) on its Najdi-Hijazi target test split fell from 55.05% to 29.96%. WER across the full SADA test set declined from 58.84% to 35.61%. WER measures substitutions, deletions and insertions relative to the number of words; it is not equivalent to saying that the model became ‘25% more accurate.’ In the same experiment, FLEURS English WER changed from 11.04% to 10.42%, and Arabic WER from 12.67% to 11.41%. These figures belong to separate test sets and should not be conflated.\n\nThe training recipe mixes target-dialect speech with weighted English and Arabic examples from FLEURS and groups utterances of similar duration to manage training computation. The best reported SADA result followed 12,000 training steps and about 4.5 hours on two GPUs. Those are NVIDIA’s experimental conditions; they do not guarantee the same cost or result with different data, model versions or infrastructure.\n\nThe post stresses that results vary by evaluation set and that a gain on one test cannot be generalized automatically to other dialects or real-world settings. Curation choices, background noise, microphones, speaking rate and transcript conventions can all affect WER. NVIDIA also reports that offline decoding options lowered WER while adding latency; a wider attention context can add about 800 milliseconds. Live captioning, call-archive transcription and meeting records therefore involve different trade-offs.\n\nSpeech recognition adapted to Saudi dialects could be a technical component for customer service, media archives or voice interfaces. The source does not say that this model has been deployed by Saudi businesses or won a commercial contract. Its measurements show how the output changed on the specified test sets under the published recipe. Real deployments would still need local user testing, error analysis, privacy and consent controls, accessibility review, operating-cost assessment and appropriate human oversight.\n\nThe table reports WER percentages and keeps test sets separate. Before-and-after values are the measurements in NVIDIA’s technical post. They are experimental results and do not claim to represent every dialect or speaker in the country. The supported conclusion is that dialect-targeted fine-tuning improved the measured test performance in this experiment; performance in actual use requires separate validation.","corrections":[],"editorial_illustration":{"kind":"symbolic-editorial-illustration","documentary":false,"scope":"section","asset_id":"business","url":"https://haberpaper.com/assets/illustrations/business-1536.webp","width":1536,"height":1024,"alt":"Symbolic industrial workshop with machinery, a warehouse and a crane.","language":"en"},"section_illustration":{"kind":"symbolic-editorial-illustration","documentary":false,"scope":"section","asset_id":"business","url":"https://haberpaper.com/assets/illustrations/business-1536.webp","width":1536,"height":1024,"alt":"Symbolic industrial workshop with machinery, a warehouse and a crane.","language":"en"}}}