--- title: Breaking language barriers 2.0: Moving closer towards... description: Whisper-Hindi v2 is a fine-tuned Hindi ASR model built on OpenAI's Whisper, trained on 3,000 hours of cleaned Hindi speech data. Key improvements over v1... canonical: https://daily.dev/posts/breaking-language-barriers-2-0-moving-closer-towards-fully-reliable-production-ready-hindi-asr-ayjehj8a9 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Breaking language barriers 2.0: Moving closer towards fully reliable, production-ready Hindi ASR | daily.dev og:description: Whisper-Hindi v2 is a fine-tuned Hindi ASR model built on OpenAI's Whisper, trained on 3,000 hours of cleaned Hindi speech data. Key improvements over v1... og:url: https://daily.dev/posts/breaking-language-barriers-2-0-moving-closer-towards-fully-reliable-production-ready-hindi-asr-ayjehj8a9 og:image: https://api.daily.dev/og/posts/AYjEHJ8A9.png og:image:alt: Breaking language barriers 2.0: Moving closer towards fully reliable, production-ready Hindi ASR 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. # Breaking language barriers 2.0: Moving closer towards fully reliable, production-ready Hindi ASR **[Collabora Blog](https://daily.dev/sources/collabora)** · 6 min read · 0 upvotes · 0 comments ## Summary Whisper-Hindi v2 is a fine-tuned Hindi ASR model built on OpenAI's Whisper, trained on 3,000 hours of cleaned Hindi speech data. Key improvements over v1 include explicit timestamp prediction, a WebDataset-based pipeline delivering 5–6× faster I/O, and fine-tuning across all Whisper model sizes (Tiny through Large-v2) on a single RTX 4090. The best model achieves ~5% WER on google/fleurs under Whisper normalization, down from a baseline of ~21.5%. Indic normalization is highlighted as critical for preserving Hindi diacritics (matras) and conjuncts that standard Whisper normalization destroys. Fine-tuning reduces WER by over 15× on Tiny/Base models and 4–5× on Medium/Large-v2. Training used mixed-precision, 8-bit AdamW for Large-v2, and carefully tuned learning rates well below the original Whisper paper values. Models are available on Hugging Face. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What WER does fine-tuned Whisper Large-v2 achieve on Hindi using google/fleurs? Fine-tuned Whisper Large-v2 achieves 5.33% WER under Whisper normalization and 13.06% under Indic normalization on google/fleurs, compared to a baseline of 21.54% and 38.46% respectively before fine-tuning. The model was trained on 3,000 hours of Hindi speech on a single NVIDIA RTX 4090 using mixed-precision training and 8-bit AdamW. _Teams building Hindi speech pipelines track benchmark results like these on daily.dev._ ### Why does standard Whisper normalization break Hindi text and what is the alternative? Standard Whisper normalization splits conjuncts like क्ष and strips aspirate matras like ढ़, turning valid Hindi words into unreadable text. Indic normalization preserves every matra and conjunct cluster, maintaining semantic integrity and readability. This matters for downstream NLP tasks, subtitles, and live transcription where missing diacritics break both comprehension and model accuracy. _Developers working on Indic language NLP find normalization edge cases like this discussed on daily.dev._ ### How much faster is WebDataset compared to standard file-based I/O for Whisper fine-tuning? Switching to WebDataset shards for precomputed log-Mel arrays and token ID sequences with timestamp markers delivered 5–6× faster training throughput compared to reading millions of individual audio files from disk. The key is eliminating per-step file system overhead by packing all preprocessed data into sequential shards read during training. _Engineers optimizing large-scale audio training pipelines share throughput wins 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 --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#deep-learning](https://daily.dev/tags/deep-learning), [#nlp](https://daily.dev/tags/nlp), [#whisper](https://daily.dev/tags/whisper) [View this post on daily.dev](https://daily.dev/posts/breaking-language-barriers-2-0-moving-closer-towards-fully-reliable-production-ready-hindi-asr-ayjehj8a9) ```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":"Breaking language barriers 2.0: Moving closer towards fully reliable, production-ready Hindi ASR","url":"https://daily.dev/posts/breaking-language-barriers-2-0-moving-closer-towards-fully-reliable-production-ready-hindi-asr-ayjehj8a9","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/breaking-language-barriers-2-0-moving-closer-towards-fully-reliable-production-ready-hindi-asr-ayjehj8a9"},"datePublished":"2026-08-11T08:54:25.827Z","dateModified":"2026-08-11T09:28:19.485Z","description":"Whisper-Hindi v2 is a fine-tuned Hindi ASR model built on OpenAI's Whisper, trained on 3,000 hours of cleaned Hindi speech data. Key improvements over v1...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/031545dbbb22d4ebf8c84136dd7d9a7a?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/031545dbbb22d4ebf8c84136dd7d9a7a?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Collabora Blog","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":"Collabora Blog","logo":"https://media.daily.dev/image/upload/s--ZGZ0FjqS--/f_auto,q_auto/v1786438306/logos/collabora?_a=BAMAMicg0","url":"https://daily.dev/sources/collabora"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/breaking-language-barriers-2-0-moving-closer-towards-fully-reliable-production-ready-hindi-asr-ayjehj8a9","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,deep-learning,nlp,whisper","timeRequired":"PT6M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collabora Blog","item":"https://daily.dev/sources/collabora"},{"@type":"ListItem","position":3,"name":"Breaking language barriers 2.0: Moving closer towards fully reliable, production-ready Hindi ASR"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/breaking-language-barriers-2-0-moving-closer-towards-fully-reliable-production-ready-hindi-asr-ayjehj8a9#faq","mainEntity":[{"@type":"Question","name":"What WER does fine-tuned Whisper Large-v2 achieve on Hindi using google/fleurs?","acceptedAnswer":{"@type":"Answer","text":"Fine-tuned Whisper Large-v2 achieves 5.33% WER under Whisper normalization and 13.06% under Indic normalization on google/fleurs, compared to a baseline of 21.54% and 38.46% respectively before fine-tuning. The model was trained on 3,000 hours of Hindi speech on a single NVIDIA RTX 4090 using mixed-precision training and 8-bit AdamW. Teams building Hindi speech pipelines track benchmark results like these on daily.dev."}},{"@type":"Question","name":"Why does standard Whisper normalization break Hindi text and what is the alternative?","acceptedAnswer":{"@type":"Answer","text":"Standard Whisper normalization splits conjuncts like क्ष and strips aspirate matras like ढ़, turning valid Hindi words into unreadable text. Indic normalization preserves every matra and conjunct cluster, maintaining semantic integrity and readability. This matters for downstream NLP tasks, subtitles, and live transcription where missing diacritics break both comprehension and model accuracy. Developers working on Indic language NLP find normalization edge cases like this discussed on daily.dev."}},{"@type":"Question","name":"How much faster is WebDataset compared to standard file-based I/O for Whisper fine-tuning?","acceptedAnswer":{"@type":"Answer","text":"Switching to WebDataset shards for precomputed log-Mel arrays and token ID sequences with timestamp markers delivered 5–6× faster training throughput compared to reading millions of individual audio files from disk. The key is eliminating per-step file system overhead by packing all preprocessed data into sequential shards read during training. Engineers optimizing large-scale audio training pipelines share throughput wins like this on daily.dev."}}]} ```