--- title: Fine-Tuning NVIDIA Nemotron for Saudi Arabic Dialects,... description: A technical walkthrough details how NVIDIA Nemotron 3.5 ASR was fine-tuned for Najdi and Hijazi Saudi Arabic dialects using the NeMo framework. Key techniques... canonical: https://daily.dev/posts/fine-tuning-nvidia-nemotron-for-saudi-arabic-dialects-with-a-path-to-other-languages-6uqddopfl twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Fine-Tuning NVIDIA Nemotron for Saudi Arabic Dialects, with a Path to Other Languages | daily.dev og:description: A technical walkthrough details how NVIDIA Nemotron 3.5 ASR was fine-tuned for Najdi and Hijazi Saudi Arabic dialects using the NeMo framework. Key techniques... og:url: https://daily.dev/posts/fine-tuning-nvidia-nemotron-for-saudi-arabic-dialects-with-a-path-to-other-languages-6uqddopfl og:image: https://api.daily.dev/og/posts/6uQddopfL.png og:image:alt: Fine-Tuning NVIDIA Nemotron for Saudi Arabic Dialects, with a Path to Other Languages og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Fine-Tuning NVIDIA Nemotron for Saudi Arabic Dialects, with a Path to Other Languages **[NVIDIA Developer](https://daily.dev/sources/nvidiadev)** · 14 min read · 0 upvotes · 0 comments ## Summary A technical walkthrough details how NVIDIA Nemotron 3.5 ASR was fine-tuned for Najdi and Hijazi Saudi Arabic dialects using the NeMo framework. Key techniques include minimal data curation (removing unusable annotations rather than hard accents), replay mixing (10% FLEURS data to prevent catastrophic forgetting of English/MSA), duration-based bucketing for efficient training, and testing different encoder-unfreezing depths. Results show WER dropped from 55.05% to 29.96% on target dialects, with English and Arabic retention metrics actually improving slightly. Further gains came from inference-time tweaks: widening attention context reduced WER by 1.31 points without retraining, and beam search (MALSD beam-8) pushed WER down to 27.25%. The piece closes by generalizing the recipe to other languages and mentions the newly released Nemotron 3 Diarization model for speaker-attributed transcripts. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How much does replay mixing with FLEURS data help prevent catastrophic forgetting when fine-tuning an ASR model on a new dialect? Mixing in just 7% English and 3% Arabic FLEURS data against 90% target dialect speech not only prevented forgetting but slightly improved retention: FLEURS English WER improved from 11.04% to 10.42% and FLEURS Arabic WER from 12.67% to 11.41%, while SADA Najdi and Hijazi WER dropped from 55.05% to 29.96% after fine-tuning. _daily.dev surfaces practical fine-tuning tricks like this for teams adapting ASR models to new languages._ ### Does unfreezing more encoder layers always give better accuracy when fine-tuning a streaming ASR model like Nemotron 3.5? Yes, in this case more trainable capacity consistently improved results: updating only the top 6 encoder layers gave 33.42% WER, top 8 layers gave 32.32%, and unfreezing all 24 layers gave the best result at 29.96% WER, a 2.4 point gain over partial unfreezing, with 134 hours of target speech data available. Partial freezing remains preferable when data is scarcer or memory-constrained. _developers tuning adaptation depth for their own ASR datasets can find comparisons like this on daily.dev._ ### Can I improve ASR transcription accuracy without retraining the model, just by changing decoding settings? Yes, switching the Nemotron 3.5 ASR streaming attention context from the default [56,3] to the widest supported [56,13] reduced WER by 1.31 absolute points with no retraining, at the cost of about 800ms additional latency. Combining that with MALSD beam-8 decoding pushed WER down further to 27.25%, compared to 29.96% with greedy decoding at the default context. _anyone tuning inference-time latency versus accuracy trade-offs can track techniques like this on daily.dev._ ## Similar posts on daily.dev - [How to Fine-Tune Nemotron 3.5 ASR for Your Language, Domain, or Accent](https://daily.dev/posts/how-to-fine-tune-nemotron-3-5-asr-for-your-language-domain-or-accent-lmayxxp44) · Hugging Face · 0 upvotes · 0 comments - [Evaluate Clinical ASR Models Faster with Agent Skills and NVIDIA Nemotron Speech](https://daily.dev/posts/evaluate-clinical-asr-models-faster-with-agent-skills-and-nvidia-nemotron-speech-zrlobjd2v) · NVIDIA Developer · 1 upvotes · 0 comments --- Tags: [#deep-learning](https://daily.dev/tags/deep-learning), [#nvidia](https://daily.dev/tags/nvidia), [#speech-recognition](https://daily.dev/tags/speech-recognition) [View this post on daily.dev](https://daily.dev/posts/fine-tuning-nvidia-nemotron-for-saudi-arabic-dialects-with-a-path-to-other-languages-6uqddopfl) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Fine-Tuning NVIDIA Nemotron for Saudi Arabic Dialects, with a Path to Other Languages","url":"https://daily.dev/posts/fine-tuning-nvidia-nemotron-for-saudi-arabic-dialects-with-a-path-to-other-languages-6uqddopfl","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/fine-tuning-nvidia-nemotron-for-saudi-arabic-dialects-with-a-path-to-other-languages-6uqddopfl"},"datePublished":"2026-10-01T05:01:38.572Z","dateModified":"2026-10-01T05:02:02.597Z","description":"A technical walkthrough details how NVIDIA Nemotron 3.5 ASR was fine-tuned for Najdi and Hijazi Saudi Arabic dialects using the NeMo framework. Key techniques...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/0241fce9b366e863af92800e04508b07?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/0241fce9b366e863af92800e04508b07?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"NVIDIA Developer","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"NVIDIA Developer","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/86e45aab42ba48ce83103d01b1119910","url":"https://daily.dev/sources/nvidiadev"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/fine-tuning-nvidia-nemotron-for-saudi-arabic-dialects-with-a-path-to-other-languages-6uqddopfl","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"deep-learning,nvidia,speech-recognition","timeRequired":"PT14M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"NVIDIA Developer","item":"https://daily.dev/sources/nvidiadev"},{"@type":"ListItem","position":3,"name":"Fine-Tuning NVIDIA Nemotron for Saudi Arabic Dialects, with a Path to Other Languages"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/fine-tuning-nvidia-nemotron-for-saudi-arabic-dialects-with-a-path-to-other-languages-6uqddopfl#faq","mainEntity":[{"@type":"Question","name":"How much does replay mixing with FLEURS data help prevent catastrophic forgetting when fine-tuning an ASR model on a new dialect?","acceptedAnswer":{"@type":"Answer","text":"Mixing in just 7% English and 3% Arabic FLEURS data against 90% target dialect speech not only prevented forgetting but slightly improved retention: FLEURS English WER improved from 11.04% to 10.42% and FLEURS Arabic WER from 12.67% to 11.41%, while SADA Najdi and Hijazi WER dropped from 55.05% to 29.96% after fine-tuning. daily.dev surfaces practical fine-tuning tricks like this for teams adapting ASR models to new languages."}},{"@type":"Question","name":"Does unfreezing more encoder layers always give better accuracy when fine-tuning a streaming ASR model like Nemotron 3.5?","acceptedAnswer":{"@type":"Answer","text":"Yes, in this case more trainable capacity consistently improved results: updating only the top 6 encoder layers gave 33.42% WER, top 8 layers gave 32.32%, and unfreezing all 24 layers gave the best result at 29.96% WER, a 2.4 point gain over partial unfreezing, with 134 hours of target speech data available. Partial freezing remains preferable when data is scarcer or memory-constrained. developers tuning adaptation depth for their own ASR datasets can find comparisons like this on daily.dev."}},{"@type":"Question","name":"Can I improve ASR transcription accuracy without retraining the model, just by changing decoding settings?","acceptedAnswer":{"@type":"Answer","text":"Yes, switching the Nemotron 3.5 ASR streaming attention context from the default [56,3] to the widest supported [56,13] reduced WER by 1.31 absolute points with no retraining, at the cost of about 800ms additional latency. Combining that with MALSD beam-8 decoding pushed WER down further to 27.25%, compared to 29.96% with greedy decoding at the default context. anyone tuning inference-time latency versus accuracy trade-offs can track techniques like this on daily.dev."}}]} ```