{"id":170,"date":"2026-07-06T12:55:20","date_gmt":"2026-07-06T12:55:20","guid":{"rendered":"https:\/\/datahive.ai\/blog\/?p=170"},"modified":"2026-07-13T11:46:15","modified_gmt":"2026-07-13T11:46:15","slug":"the-data-moat-in-dialectal-arabic-speech-recognition","status":"publish","type":"post","link":"https:\/\/datahive.ai\/blog\/2026\/07\/06\/the-data-moat-in-dialectal-arabic-speech-recognition\/","title":{"rendered":"The Data Moat in Dialectal Arabic Speech Recognition"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\"><em>Why off-the-shelf \u2014 and even frontier \u2014 ASR fails on Arabic dialects, and what a targeted corpus fixes.<\/em><\/p><div id=\"ez-toc-container\" class=\"ez-toc-v2_0_85 ez-toc-wrap-left counter-hierarchy ez-toc-counter ez-toc-black ez-toc-container-direction\">\n<div class=\"ez-toc-title-container\">\n<p class=\"ez-toc-title\" style=\"cursor:inherit\">Table of Contents<\/p>\n<span class=\"ez-toc-title-toggle\"><a href=\"#\" class=\"ez-toc-pull-right ez-toc-btn ez-toc-btn-xs ez-toc-btn-default ez-toc-toggle\" aria-label=\"Toggle Table of Content\"><span class=\"ez-toc-js-icon-con\"><span class=\"\"><span class=\"eztoc-hide\" style=\"display:none;\">Toggle<\/span><span class=\"ez-toc-icon-toggle-span\"><svg style=\"fill: #999999;color:#999999\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" class=\"list-377408\" width=\"20px\" height=\"20px\" viewBox=\"0 0 24 24\" fill=\"none\"><path d=\"M6 6H4v2h2V6zm14 0H8v2h12V6zM4 11h2v2H4v-2zm16 0H8v2h12v-2zM4 16h2v2H4v-2zm16 0H8v2h12v-2z\" fill=\"currentColor\"><\/path><\/svg><svg style=\"fill: #999999;color:#999999\" class=\"arrow-unsorted-368013\" xmlns=\"http:\/\/www.w3.org\/2000\/svg\" width=\"10px\" height=\"10px\" viewBox=\"0 0 24 24\" version=\"1.2\" baseProfile=\"tiny\"><path d=\"M18.2 9.3l-6.2-6.3-6.2 6.3c-.2.2-.3.4-.3.7s.1.5.3.7c.2.2.4.3.7.3h11c.3 0 .5-.1.7-.3.2-.2.3-.5.3-.7s-.1-.5-.3-.7zM5.8 14.7l6.2 6.3 6.2-6.3c.2-.2.3-.5.3-.7s-.1-.5-.3-.7c-.2-.2-.4-.3-.7-.3h-11c-.3 0-.5.1-.7.3-.2.2-.3.5-.3.7s.1.5.3.7z\"\/><\/svg><\/span><\/span><\/span><\/a><\/span><\/div>\n<nav><ul class='ez-toc-list ez-toc-list-level-1 eztoc-toggle-hide-by-default' ><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-1\" href=\"https:\/\/datahive.ai\/blog\/2026\/07\/06\/the-data-moat-in-dialectal-arabic-speech-recognition\/#The_problem_where_Arabic_ASR_breaks\" >The problem: where Arabic ASR breaks<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-2\" href=\"https:\/\/datahive.ai\/blog\/2026\/07\/06\/the-data-moat-in-dialectal-arabic-speech-recognition\/#The_corpus_built_to_test_the_hypothesis\" >The corpus: built to test the hypothesis<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-3\" href=\"https:\/\/datahive.ai\/blog\/2026\/07\/06\/the-data-moat-in-dialectal-arabic-speech-recognition\/#The_approach_its_the_data_not_the_model\" >The approach: it&#8217;s the data, not the model<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-4\" href=\"https:\/\/datahive.ai\/blog\/2026\/07\/06\/the-data-moat-in-dialectal-arabic-speech-recognition\/#Results_the_corpus_is_the_driver\" >Results: the corpus is the driver<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-5\" href=\"https:\/\/datahive.ai\/blog\/2026\/07\/06\/the-data-moat-in-dialectal-arabic-speech-recognition\/#How_we_stack_up_against_the_field\" >How we stack up against the field<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-6\" href=\"https:\/\/datahive.ai\/blog\/2026\/07\/06\/the-data-moat-in-dialectal-arabic-speech-recognition\/#How_we_measure_and_why_you_can_trust_it\" >How we measure (and why you can trust it)<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-7\" href=\"https:\/\/datahive.ai\/blog\/2026\/07\/06\/the-data-moat-in-dialectal-arabic-speech-recognition\/#What_this_means_for_you\" >What this means for you<\/a><\/li><li class='ez-toc-page-1 ez-toc-heading-level-2'><a class=\"ez-toc-link ez-toc-heading-8\" href=\"https:\/\/datahive.ai\/blog\/2026\/07\/06\/the-data-moat-in-dialectal-arabic-speech-recognition\/#Appendix_%E2%80%94_methodology_notes\" >Appendix \u2014 methodology &amp; notes<\/a><\/li><\/ul><\/nav><\/div>\n\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">Generic speech-to-text \u2014 and even much larger or paid systems \u2014 breaks down on&nbsp;<strong>dialectal, emotional Arabic<\/strong>. We show the fix is&nbsp;<strong>data, not model size<\/strong>: a targeted 4-dialect corpus, added to public data,&nbsp;<strong>roughly halves error<\/strong>&nbsp;on the covered dialects, with a controlled ablation isolating the corpus as the cause \u2014 and the gain is largest exactly where off-the-shelf models are weakest. We benchmark honestly against Meta&#8217;s 7B omniASR and the Deepgram API:&nbsp;<strong>on the speech we serve, our model leads; on generic data, frontier systems are competitive.<\/strong>&nbsp;The differentiator is the right data plus rigorous, non-cherry-picked evaluation.<\/p>\n<\/blockquote>\n\n\n\n<div class=\"wp-block-columns is-layout-flex wp-container-core-columns-is-layout-8f761849 wp-block-columns-is-layout-flex\">\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\">\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"The_problem_where_Arabic_ASR_breaks\"><\/span>The problem: where Arabic ASR breaks<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Arabic ASR is not one problem. Modern systems handle&nbsp;<strong>Modern Standard Arabic (MSA)<\/strong>&nbsp;\u2014 broadcast, formal speech \u2014 reasonably well. They fail on&nbsp;<strong>spontaneous, dialectal, emotional<\/strong>&nbsp;speech, which is most of how people actually talk. Three things make it hard:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Distance from MSA.<\/strong>&nbsp;Dialects like Moroccan Darija diverge sharply from MSA in vocabulary, morphology and phonology; MSA-trained models have little to draw on.<\/li>\n\n\n\n<li><strong>No standard orthography.<\/strong>&nbsp;Dialects have no agreed way to be&nbsp;<em>written<\/em>. The same word is spelled several ways, and clitics (\u0648\/\u0628\/\u0627\u0644) agglutinate onto words. This wrecks word-level error metrics (\u00a76).<\/li>\n\n\n\n<li><strong>Emotion shifts the acoustics.<\/strong>&nbsp;Angry, sad or happy speech changes rate, pitch and articulation \u2014 further off-distribution for models trained on neutral, read speech.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">On our held-out target test (proprietary, speaker\/script\/text-disjoint), stock Whisper-large-v3 scores&nbsp;<strong>WER 0.589 \/ CER 0.226<\/strong>&nbsp;\u2014 unusable for production. On Moroccan Darija it is far worse.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"The_corpus_built_to_test_the_hypothesis\"><\/span>The corpus: built to test the hypothesis<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">To find out whether&nbsp;<strong>targeted data \u2014 not a bigger model \u2014 is what closes the gap<\/strong>, we needed speech that off-the-shelf systems have never seen and that public datasets simply do not cover: dialectal, emotional, conversational Arabic. So we built it ourselves. Through&nbsp;<strong>DataHive&#8217;s data-collection platform<\/strong>&nbsp;we recruited native speakers of each target dialect and ran a structured recording pipeline \u2014 scripted prompts across ten everyday domains (customer service, healthcare, finance, education, travel, emergency, family, daily life, shopping, technology), each performed in four emotional registers, with multiple takes and per-clip quality control.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The result is our&nbsp;<strong>proprietary 4-dialect \u00d7 4-emotion corpus<\/strong>&nbsp;\u2014 native-speaker recordings balanced across emotions and&nbsp;<strong>ten everyday domains<\/strong>, spanning&nbsp;<strong>11,000+ distinct scripts<\/strong>, with&nbsp;<strong>speaker-, script-, and text-disjoint<\/strong>&nbsp;splits so the numbers later in this paper measure genuine generalization to unseen speakers and sentences, not memorization. And because it comes from a&nbsp;<strong>repeatable collection platform<\/strong>&nbsp;rather than a one-off dataset, it is an asset we extend to new dialects, domains, and emotions on demand \u2014 the coverage grows to fit the use case.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"The_approach_its_the_data_not_the_model\"><\/span>The approach: it&#8217;s the data, not the model<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">We fine-tune Whisper-large-v3 with LoRA on a data mixture, and run a&nbsp;<strong>controlled ablation<\/strong>&nbsp;to isolate what drives quality. Two recipes differ by exactly one thing:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Recipe A \u2014 public only:<\/strong>&nbsp;open Arabic datasets (SADA + MASC).<\/li>\n\n\n\n<li><strong>Recipe B \u2014 public + ours:<\/strong>&nbsp;the same public data&nbsp;<strong>plus<\/strong>&nbsp;our proprietary 4-dialect \u00d7 4-emotion corpus.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">The public mixture is held byte-identical between A and B, so any A\u2192B difference is&nbsp;<strong>our corpus, nothing else<\/strong>. Every model is scored on the&nbsp;<strong>same held-out target test<\/strong>&nbsp;\u2014 only the&nbsp;<em>training<\/em>&nbsp;data changes.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"Results_the_corpus_is_the_driver\"><\/span>Results: the corpus is the driver<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Headline, target test, leaderboard normalizer:<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><thead><tr><th class=\"has-text-align-left\" data-align=\"left\">Model<\/th><th class=\"has-text-align-right\" data-align=\"right\">WER<\/th><th class=\"has-text-align-right\" data-align=\"right\">CER<\/th><\/tr><\/thead><tbody><tr><th class=\"has-text-align-left\" data-align=\"left\">Stock Whisper-large-v3<\/th><td class=\"has-text-align-right\" data-align=\"right\">0.589<\/td><td class=\"has-text-align-right\" data-align=\"right\">0.226<\/td><\/tr><tr><th class=\"has-text-align-left\" data-align=\"left\">Public-only fine-tune (A)<\/th><td class=\"has-text-align-right\" data-align=\"right\">0.598<\/td><td class=\"has-text-align-right\" data-align=\"right\">0.204<\/td><\/tr><tr><th class=\"has-text-align-left\" data-align=\"left\">+ our corpus (B)<\/th><td class=\"has-text-align-right\" data-align=\"right\">0.258<\/td><td class=\"has-text-align-right\" data-align=\"right\">0.077<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Public data alone does not solve the dialectal problem<\/strong>&nbsp;\u2014 adding more public Arabic (A) leaves word error at stock and only trims character error modestly.&nbsp;<strong>On the target domain the corpus is what moves the needle:<\/strong>&nbsp;B cuts WER by 56% and CER by 66% versus stock, and the corpus contribution itself \u2014&nbsp;<strong>B \u2212 A<\/strong>, public mix held identical \u2014 is large and significant on every dialect (overall \u22120.127 CER \/ \u22120.340 WER; \u22120.094 \/ \u22120.226 under CAMeL). This is a target-domain result \u2014 \u00a75 shows it does&nbsp;<em>not<\/em>&nbsp;hold on neutral, out-of-domain Saudi data.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The value of targeted data scales with where models fail<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"478\" src=\"https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/1-color-1-1024x478.png\" alt=\"\" class=\"wp-image-216\" srcset=\"https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/1-color-1-1024x478.png 1024w, https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/1-color-1-300x140.png 300w, https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/1-color-1-768x358.png 768w, https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/1-color-1-1536x716.png 1536w, https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/1-color-1.png 1664w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n<\/div>\n<\/div>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"wp-block-paragraph\">Target-test WER by dialect (lower = better). The corpus&#8217;s biggest wins land on Moroccan Darija \u2014 exactly where off-the-shelf collapses \u2014 and smaller wins on the MSA-adjacent Saudi dialects, where stock is bad but not catastrophic. Note the&nbsp;<em>overall<\/em>&nbsp;figures are micro-averaged and Moroccan + Najdi are 77% of the clips; the unweighted macro-average is a touch milder (stock WER 0.519 \u2192 B 0.237, \u221254%).<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"How_we_stack_up_against_the_field\"><\/span>How we stack up against the field<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">We compared our corpus-tuned model against two strong baselines on the same target test: Meta&#8217;s&nbsp;<strong>omniASR-LLM-7B<\/strong>&nbsp;(a 7-billion-parameter open model) and&nbsp;<strong>Deepgram<\/strong>&nbsp;(a commercial paid API with dialect-specific Arabic models).<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Overall WER on our target test \u2014 vs the field<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"404\" src=\"https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/2-color-1-1024x404.png\" alt=\"\" class=\"wp-image-217\" srcset=\"https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/2-color-1-1024x404.png 1024w, https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/2-color-1-300x118.png 300w, https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/2-color-1-768x303.png 768w, https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/2-color-1-1536x606.png 1536w, https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/2-color-1.png 1664w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n<\/blockquote>\n\n\n\n<p class=\"wp-block-paragraph\">Both baselines beat stock Whisper \u2014 omniASR-7B and Deepgram are genuinely strong \u2014 yet our corpus on plain Whisper beats them by a wide margin (overall \u22120.198 WER vs omniASR, significant).<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Deepgram is a competent Arabic recognizer, but on our domain the corpus-tuned model cuts its error on every dialect by roughly&nbsp;<strong>a third to a half<\/strong>&nbsp;depending on dialect and normalizer (about \u221236% overall on CER under the convention-robust metric, up to ~50% on the leaderboard metric).<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">The honest caveat: domain, not universal superiority<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">This is the speech&nbsp;<strong>we serve<\/strong>&nbsp;\u2014 our dialects, our convention. Step outside it and the wins narrow or vanish, and we report that plainly:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Neutral third-party data (Meta&#8217;s public Omnilingual).<\/strong>&nbsp;The corpus advantage concentrates the same way \u2014&nbsp;<strong>large on Moroccan<\/strong>&nbsp;(B CER 0.219 vs stock 0.398, CAMeL) \u2014 but&nbsp;<strong>near-zero on MSA-adjacent Saudi<\/strong>: under the convention-robust metric public-only training (A) matches or slightly beats the corpus (B) there, and B&#8217;s edge over stock is small.<\/li>\n\n\n\n<li><strong>External Moroccan.<\/strong>&nbsp;We beat omniASR-7B on adiren7 (YouTube Moroccan), but&nbsp;<strong>lose to it on MDER-MA \u2014 another external Moroccan set \u2014 by a wide, significant margin<\/strong>&nbsp;(B CER 0.660 vs 0.514, CAMeL), and roughly&nbsp;<strong>tie<\/strong>&nbsp;it overall on the neutral Casablanca benchmark (a shade behind on its Moroccan subset). The corpus makes Whisper better on&nbsp;<strong>our<\/strong>&nbsp;Moroccan speech, not Moroccan in general.<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">Our advantage is&nbsp;<strong>specialization for the target use case<\/strong>, not a universally better Arabic model. Saying so plainly is the point (\u00a76).<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"How_we_measure_and_why_you_can_trust_it\"><\/span>How we measure (and why you can trust it)<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Dialectal Arabic ASR is unusually easy to measure&nbsp;<em>badly<\/em>. We make three deliberate choices.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>We report CER, not just WER.<\/strong>&nbsp;Because dialects agglutinate clitics and have no standard spelling, word-level WER swings with segmentation and orthography, while character-level CER is robust. We report both and read CER as the trustworthy signal.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>We account for orthographic convention.<\/strong>&nbsp;Each system writes its own spelling. Scored against references in&nbsp;<em>our<\/em>&nbsp;convention, our model has a home advantage; against a third party&#8217;s convention it is penalized. Re-scoring under a convention-robust normalization (CAMeL-standard) shows that a sizable part of any model&#8217;s apparent lead is spelling, not quality \u2014 and the effect cuts both ways.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Under a stricter, convention-robust metric \u2014 the lead shrinks but holds<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"478\" src=\"https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/3-color-1-1024x478.png\" alt=\"\" class=\"wp-image-215\" srcset=\"https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/3-color-1-1024x478.png 1024w, https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/3-color-1-300x140.png 300w, https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/3-color-1-768x358.png 768w, https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/3-color-1-1536x716.png 1536w, https:\/\/datahive.ai\/blog\/wp-content\/uploads\/2026\/07\/3-color-1.png 1664w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Overall CER. Everyone&#8217;s number drops once spelling is folded away \u2014 but the ranking holds and B stays lowest. Our CER lead over omniASR-7B goes from \u22120.063 to \u22120.039 (still significant); over stock it barely moves (\u22120.149 \u2192 \u22120.134): that gap is real quality, not orthography.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>We don&#8217;t cherry-pick.<\/strong>&nbsp;We report the comparisons where we win&nbsp;<em>and<\/em>&nbsp;the ones where we don&#8217;t, under more than one metric and normalizer. For a buyer evaluating a vendor, that transparency is the signal: the numbers are real.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"What_this_means_for_you\"><\/span>What this means for you<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li>If your use case is&nbsp;<strong>dialectal or emotional Arabic<\/strong>, off-the-shelf and even paid\/frontier ASR will under-deliver \u2014 most on the hardest dialects.<\/li>\n\n\n\n<li><strong>Targeted data is the moat.<\/strong>&nbsp;A well-collected, on-target dialectal corpus produces gains that model size and mountains of public data do not \u2014 quality and targeting, not raw volume, move the needle.<\/li>\n\n\n\n<li>The gain is&nbsp;<strong>measurable and targeted<\/strong>: it is largest on the dialects the corpus covers, and concentrated where generic models are weakest.<\/li>\n\n\n\n<li>DataHive brings both the&nbsp;<strong>proprietary data<\/strong>&nbsp;and a&nbsp;<strong>rigorous, honest evaluation method<\/strong>&nbsp;to prove the lift on&nbsp;<em>your<\/em>&nbsp;domain.<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\"><span class=\"ez-toc-section\" id=\"Appendix_%E2%80%94_methodology_notes\"><\/span>Appendix \u2014 methodology &amp; notes<span class=\"ez-toc-section-end\"><\/span><\/h2>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Model:<\/strong>&nbsp;Whisper-large-v3 + LoRA (PEFT). Corpus share 40% of the training mixture.<\/li>\n\n\n\n<li><strong>Ablation:<\/strong>&nbsp;Recipe A redistributes the corpus slot to public data, holding the public mix byte-identical to B; the only variable is our corpus.<\/li>\n\n\n\n<li><strong>Test:<\/strong>&nbsp;held-out, speaker + script + text-disjoint; ~3,354 clips across four dialects. Headline &#8220;overall&#8221; figures are micro-averages over clips (Moroccan + Najdi are 77% of the test); we also report the 4-dialect macro-average, which is less sensitive to the speaker mix.<\/li>\n\n\n\n<li><strong>Metrics:<\/strong>&nbsp;WER + CER under the Open Universal Arabic ASR Leaderboard normalizer and a CAMeL-standard convention-robust normalizer; 95% CIs from paired bootstrap.<\/li>\n\n\n\n<li><strong>Baselines:<\/strong>&nbsp;stock Whisper-large-v3; public-only fine-tune; Meta omniASR-LLM-7B; Deepgram. External models are not scored on the Omnilingual dialect sets (omniASR&#8217;s own training data). Decoding note: Deepgram was given dialect-specific language codes (ar-SA \/ ar-JO \/ ar-MA); omniASR-7B ran through its standard fairseq2 pipeline without per-clip dialect language hints \u2014 a variable we did not fully control.<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Why off-the-shelf \u2014 and even frontier \u2014 ASR fails on Arabic dialects, and what a targeted corpus fixes. Generic speech-to-text \u2014 and even much larger or paid systems \u2014 breaks down on&nbsp;dialectal, emotional Arabic. We show the fix is&nbsp;data, not model size: a targeted 4-dialect corpus, added to public data,&nbsp;roughly halves error&nbsp;on the covered dialects, [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":222,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3],"tags":[],"class_list":["post-170","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-research"],"_links":{"self":[{"href":"https:\/\/datahive.ai\/blog\/wp-json\/wp\/v2\/posts\/170","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/datahive.ai\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/datahive.ai\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/datahive.ai\/blog\/wp-json\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/datahive.ai\/blog\/wp-json\/wp\/v2\/comments?post=170"}],"version-history":[{"count":14,"href":"https:\/\/datahive.ai\/blog\/wp-json\/wp\/v2\/posts\/170\/revisions"}],"predecessor-version":[{"id":239,"href":"https:\/\/datahive.ai\/blog\/wp-json\/wp\/v2\/posts\/170\/revisions\/239"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/datahive.ai\/blog\/wp-json\/wp\/v2\/media\/222"}],"wp:attachment":[{"href":"https:\/\/datahive.ai\/blog\/wp-json\/wp\/v2\/media?parent=170"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/datahive.ai\/blog\/wp-json\/wp\/v2\/categories?post=170"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/datahive.ai\/blog\/wp-json\/wp\/v2\/tags?post=170"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}