Dataset Viewer
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Parquet error: Scan size limit exceeded: attempted to read 1370565234 bytes, limit is 300000000 bytes Make sure that 1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files
Error code:   TooBigContentError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

🗂️ visualears-persian-asr-16k

English + فارسی · Part of Shenava 1.0 · Project hub · SLT paper submission

🌟 At a glance | معرفی سریع

English فارسی
🎯 Purpose Main public Persian ASR audio/text dataset: 3.93M 16 kHz rows. مجموعه‌دادهٔ اصلی و عمومی شنوا برای آموزش بازشناسی گفتار فارسی؛ شامل صوت ۱۶ کیلوهرتز، متن و فرادادهٔ منشأ در مقیاس چندمیلیونی.
🧩 Role flagship training corpus پیکرهٔ اصلی آموزش
📦 Snapshot 176 files; approximately 512.17 GB 176 فایل؛ حدود 512.17 GB
🧱 Packaging 158 Parquet files and 0 standalone audio files 158 فایل Parquet و 0 فایل صوتی مستقل
🗣️ Language Persian / Farsi فارسی
⚖️ License Apache License 2.0 مجوز Apache 2.0

🇬🇧 English dataset card

🧭 Dataset description

Main public Persian ASR audio/text dataset: 3.93M 16 kHz rows. It is published as one component of the Shenava-1 open Persian speech stack. The collection separates training corpora, relabels, active-learning queues, calibration sets, and held-out evaluation assets so that each repository has a clear operational role.

This card documents the repository snapshot rather than claiming that every file is an interchangeable training example. Consumers must distinguish raw media, manifests, pipeline state, labels, benchmark outputs, and model-generated annotations. Preserve source identifiers and consult retained technical notes for artifact-specific fields.

🎯 Intended uses

  • Reproduce the specific Shenava/VisualEars pipeline stage named by this repository.
  • Train, filter, calibrate, review, or evaluate Persian speech systems according to the role above.
  • Study Persian ASR data quality, acoustic conditions, text normalization, hard words, and semantic error behavior.
  • Build documented derivatives while retaining source revision, row identifiers, provenance, and transformation logs.

🚫 Uses to avoid

  • Do not merge evaluation or gold sets into training data.
  • Do not assume machine-generated labels are equivalent to human-verified ground truth.
  • Do not use uncertain, non-Persian, raw, or calibration queues as clean training data without applying their review decision.
  • Do not use speech content to infer sensitive personal attributes or to identify speakers.
  • Do not publish private information found in audio or transcripts; downstream users remain responsible for privacy review.

📁 Repository structure

Audited snapshot: 176 files, approximately 512.17 GB. Common file groups: .parquet × 158, .jsonl × 4, .nemo × 3, .model × 3, .vocab × 3, no extension × 2.

Representative or largest paths:

  • data/train-00063-of-00158.parquet
  • data/train-00127-of-00158.parquet
  • data/train-00049-of-00158.parquet
  • data/train-00116-of-00158.parquet
  • data/train-00039-of-00158.parquet

Parquet repositories should be read by shard. Audio-folder repositories may require their metadata/manifests to pair media with labels. Pipeline-state and result repositories can contain JSON, JSONL, CSV, or checkpoint artifacts rather than a conventional train table.

🚀 Loading and inspection

from datasets import load_dataset

repo_id = "Reza2kn/visualears-persian-asr-16k"
dataset = load_dataset(repo_id, split="train", streaming=True)
first = next(iter(dataset))
print(first.keys())

If the repository is an artifact archive rather than a directly loadable dataset script, use snapshot_download(repo_id=repo_id, repo_type="dataset") and read its manifests or media files explicitly. Always inspect features, split names, and the card revision before launching a long job.

For audio, confirm the decoded sampling rate and channel count rather than trusting filename extensions. For text, preserve Persian Unicode, ZWNJ/half-space, digits, and original text before applying a normalized view. Never overwrite the source transcript during normalization.

🧬 Provenance and processing

The repository name and collection note define its stage in the Shenava pipeline. Derived labels, confidence signals, curriculum decisions, and quality buckets should be treated as versioned annotations. When creating a derivative, record the source repository revision, selection rule, normalization version, model revision used for pseudo-labeling, and human-review status.

✅ Quality checks

At minimum, validate file readability, unique/stable identifiers, audio duration, transcript emptiness, language, sampling rate, clipping, duplicated content, and manifest/media correspondence. For evaluation data, verify that reference text and sample membership have not changed. For generated labels, manually inspect stratified samples across confidence, duration, source, and acoustic condition.

⚠️ Biases and limitations

The data may unevenly represent accents, regions, ages, genders, speaking styles, recording devices, noise conditions, and formal versus conversational Persian. Filtering can improve average label quality while removing difficult but important speakers. Pseudo-labels inherit model errors. Literary recitation, telephony, far-field audio, and spontaneous speech are distinct domains and should not be treated as interchangeable.

🏛️ Flagship corpus facts

  • 3,927,733 rows in the public train split.
  • 512.17 GB of repository content at the audited revision.
  • 158 Parquet shards, designed for streaming and sharded processing.
  • Audio is embedded in Parquet and declared at 16,000 Hz.
  • Verified schema from the first production shard: audio (bytes, path), id (string), duration (float), and source (string).
  • The first shard contains 25,000 rows; consumers should not assume every final shard has the same row count.

The hosted viewer can exceed its scan-size limit because a row group contains embedded audio. This is a viewer constraint, not evidence of corrupt data. Prefer streaming iteration, explicit shard selection, or a local cache with enough storage. Do not clone the complete 512 GB repository accidentally during a quick experiment.

🧱 Recommended large-scale access pattern

  1. Start with streaming and validate fields on a small prefix.
  2. Apply deterministic duration, language, and quality filters before materializing data.
  3. Cache only selected shards or transformed examples.
  4. Preserve id and source through every transformation.
  5. Write a manifest containing source revision, filter code, counts, and rejected-row statistics.
  6. Keep evaluation repositories such as Golden6669 completely outside training and adaptation.

🇮🇷 کارت مجموعه‌داده به فارسی

🧭 معرفی مجموعه‌داده

مجموعه‌دادهٔ اصلی و عمومی شنوا برای آموزش بازشناسی گفتار فارسی؛ شامل صوت ۱۶ کیلوهرتز، متن و فرادادهٔ منشأ در مقیاس چندمیلیونی. این مخزن بخشی از زنجیرهٔ باز و فارسی‌محور Shenava-1 است. مجموعه، پیکرهٔ آموزش، بازبرچسب‌گذاری، صف یادگیری فعال، دادهٔ کالیبراسیون و معیارهای نگه‌داشته‌شده را جدا می‌کند تا نقش هر مخزن روشن بماند.

این کارت وضعیت یک نسخهٔ مشخص از مخزن را توضیح می‌دهد و ادعا نمی‌کند که همهٔ فایل‌ها نمونهٔ آموزشی هم‌نوع هستند. رسانهٔ خام، manifest، وضعیت خط لوله، برچسب، خروجی بنچمارک و annotation تولیدشده با مدل باید بر اساس نقش واقعی خود مصرف شوند.

🎯 کاربردهای پیشنهادی

  • بازتولید همان مرحله از خط لولهٔ Shenava/VisualEars که در نام و توضیح مخزن آمده است.
  • آموزش، پالایش، کالیبراسیون، بازبینی یا ارزیابی سامانهٔ گفتار فارسی متناسب با نقش داده.
  • مطالعهٔ کیفیت داده، شرایط صوتی، نرمال‌سازی متن، واژه‌های دشوار و خطاهای معنایی.
  • ساخت مشتق مستند همراه با حفظ نسخهٔ مبدأ، شناسهٔ ردیف، منشأ و گزارش تبدیل.

🚫 کاربردهای نامناسب

  • دادهٔ ارزیابی یا طلایی را وارد آموزش نکنید.
  • برچسب تولیدشده با مدل را بدون بازبینی معادل حقیقت انسانی ندانید.
  • صف نامطمئن، غیرفارسی، خام یا کالیبراسیون را پیش از اعمال تصمیم بازبینی، دادهٔ پاک فرض نکنید.
  • از محتوا برای شناسایی گوینده یا استنباط ویژگی حساس استفاده نکنید.
  • اطلاعات خصوصی احتمالی در صوت یا متن را بازنشر نکنید و ارزیابی حریم خصوصی را انجام دهید.

📁 ساختار مخزن

نسخهٔ ممیزی‌شده شامل 176 فایل با حجم تقریبی 512.17 GB است. گروه‌های رایج فایل: .parquet × 158, .jsonl × 4, .nemo × 3, .model × 3, .vocab × 3, no extension × 2.

مسیرهای شاخص یا بزرگ:

  • data/train-00063-of-00158.parquet
  • data/train-00127-of-00158.parquet
  • data/train-00049-of-00158.parquet
  • data/train-00116-of-00158.parquet
  • data/train-00039-of-00158.parquet

مخزن Parquet را به‌صورت شاردی بخوانید. در مخزن audio-folder، manifest یا فراداده برای اتصال صوت و برچسب لازم است. مخزن وضعیت خط لوله یا نتایج ممکن است به‌جای جدول train شامل JSON، JSONL، CSV یا چک‌پوینت باشد.

🚀 بارگذاری و بررسی

در بارگذاری با datasets ابتدا نام split و features را بررسی کنید. اگر مخزن آرشیو مصنوع است، با snapshot_download فایل‌های لازم را انتخابی دریافت کنید. نرخ نمونه‌برداری و تعداد کانال را پس از decode تأیید کنید. در متن، نسخهٔ اصلی فارسی، نیم‌فاصله و رقم‌ها را نگه دارید و نمای نرمال‌شده را در ستونی جدا بسازید.

🧬 منشأ و پردازش

نام مخزن و توضیح مجموعه، جایگاه داده را در خط لوله مشخص می‌کند. برچسب مشتق، confidence، تصمیم curriculum و سطل کیفیت را annotation نسخه‌دار بدانید. در ساخت مشتق، revision مبدأ، قاعدهٔ انتخاب، نسخهٔ نرمال‌سازی، مدل تولیدکنندهٔ pseudo-label و وضعیت بازبینی انسانی را ثبت کنید.

✅ کنترل کیفیت

خوانایی فایل، یکتایی شناسه، مدت صوت، متن خالی، زبان، نرخ نمونه، clipping، تکرار محتوا و تطابق manifest با رسانه را بررسی کنید. در دادهٔ ارزیابی، ثابت‌بودن اعضا و متن مرجع ضروری است. برای برچسب ماشینی، نمونه‌گیری لایه‌بندی‌شده بر اساس confidence، مدت، منشأ و شرایط صوتی انجام دهید.

⚠️ سوگیری‌ها و محدودیت‌ها

ممکن است لهجه، منطقه، سن، جنسیت، سبک گفتار، دستگاه ضبط، نویز و گفتار رسمی/محاوره‌ای به‌طور نامتوازن پوشش داده شوند. پالایش شدید می‌تواند میانگین کیفیت را بالا ببرد ولی گویندگان دشوار و مهم را حذف کند. pseudo-label خطای مدل را به ارث می‌برد و حوزه‌هایی مانند شعرخوانی، تلفن، دورمیدان و گفتار خودانگیخته قابل جایگزینی با یکدیگر نیستند.

🏛️ مشخصات پیکرهٔ اصلی

  • ۳٬۹۲۷٬۷۳۳ ردیف در بخش عمومی train.
  • ۵۱۲٫۱۷ گیگابایت محتوای مخزن در نسخهٔ ممیزی‌شده.
  • ۱۵۸ شارد Parquet برای پردازش جریانی و توزیع‌شده.
  • صوت داخل Parquet ذخیره شده و نرخ نمونه‌برداری آن ۱۶٬۰۰۰ هرتز است.
  • شمای تأییدشده از شارد تولیدی نخست: audio شامل bytes و path، سپس id، duration و source.
  • شارد نخست ۲۵ هزار ردیف دارد؛ نباید فرض کرد شارد پایانی دقیقاً همین تعداد را دارد.

نمایشگر میزبانی‌شده ممکن است به‌دلیل وجود صوت تعبیه‌شده در یک row group از سقف اسکن عبور کند. این محدودیت نمایشگر است و نشانهٔ خرابی داده نیست. برای آزمایش کوتاه، مخزن ۵۱۲ گیگابایتی را ناخواسته کامل clone نکنید؛ streaming، انتخاب شارد و کش کنترل‌شده مناسب‌تر است.

🧱 الگوی پیشنهادی دسترسی در مقیاس بزرگ

۱. کار را با streaming و بررسی چند نمونه آغاز کنید. ۲. پیش از ذخیره‌سازی، فیلتر مدت، زبان و کیفیت را قطعی اجرا کنید. ۳. فقط شاردها یا نمونه‌های انتخاب‌شده را کش کنید. ۴. فیلدهای id و source را در تمام تبدیل‌ها حفظ کنید. ۵. نسخهٔ مبدأ، کد فیلتر، شمارش پذیرش و رد را در manifest ثبت کنید. ۶. داده‌های ارزیابی مانند Golden6669 را کاملاً بیرون از آموزش نگه دارید.

🔬 Reproducibility | بازتولیدپذیری

  • Pin the dataset revision and record selected files or shards. | نسخهٔ مخزن و شاردهای انتخابی را ثبت کنید.
  • Keep immutable IDs and provenance columns. | شناسهٔ پایدار و ستون منشأ را حفظ کنید.
  • Publish accepted/rejected counts for every filter. | شمارش پذیرش و رد هر فیلتر را منتشر کنید.
  • Separate original, normalized, and model-generated text. | متن اصلی، نرمال‌شده و تولیدشده با مدل را جدا نگه دارید.
  • Audit train/eval overlap by content, speaker when available, and source. | هم‌پوشانی آموزش و ارزیابی را بر اساس محتوا، گویندهٔ موجود و منشأ بررسی کنید.

📚 Citation, links, and license | استناد، پیوندها و مجوز

@dataset{shenava1_visualears_persian_asr_16k,
  title  = {visualears-persian-asr-16k: a Shenava-1 Persian data artifact},
  author = {Reza2kn},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Reza2kn/visualears-persian-asr-16k}
}

Released under the Apache License 2.0. این مخزن با مجوز Apache 2.0 منتشر شده است.

Downloads last month
531

Models trained or fine-tuned on Reza2kn/visualears-persian-asr-16k

Collection including Reza2kn/visualears-persian-asr-16k