zuhri025/urdu-dacvae

---
language:
  - ur
  - en
license: cc-by-4.0
task_categories:
  - automatic-speech-recognition
tags:
  - urdu
  - speech
  - audio
  - asr
  - tts
size_categories:
  - 10K<n<100K
---

# urdu-dacvae

A cleaned, normalised Urdu (+ limited English) speech dataset derived from
multiple sources, intended for ASR / TTS / VAE latent modelling.

## Dataset Summary

| Field | Value |
|---|---|
| **Samples** | 182,472 |
| **Total audio** | 445.1 hours |
| **Duration filter** | 2.0s < duration < 20.0s |
| **Unique characters** | 91 |
| **Created** | 2026-05-19 |

## Sources

- [zuhri025/munch-1-latent-NEW-parquet](https://huggingface.co/datasets/zuhri025/munch-1-latent-NEW-parquet)

• zuhri025/IndicVoice-latent-NEW-parquet

  Text Normalisation

  All transcripts were processed with the following rules:

  Kept

  • Urdu alphabet — full Naskh character set including ں ڈ ڑ ٹ پ چ ژ گ ہ ھ ے ئ
  • Arabic diacritics / harakat — zabr (َ), zeer (ِ), pesh (ُ), shadda (ّ), sukun (ْ), tanwin variants, maddah, superscript alef (ٰ), and the full U+064B–U+065F block
  • Tatweel / kashida ـ (U+0640) — preserved for connected script rendering
  • Urdu punctuation — ۔ ، ؟
  • English A–Z / a–z and basic ASCII punctuation . , ! ? : ; - ( ) " '

  Normalised (non-canonical → canonical Urdu)

  From	To	Reason
  ك U+0643	ک U+06A9	Arabic kaf → Urdu kaf
  ي U+064A	ی U+06CC	Arabic ya → Urdu ya
  ى U+0649	ی U+06CC	Alef maqsura → Urdu ya
  ه U+0647	ہ U+06C1	Arabic ha → Urdu gol-he
  أ إ ٱ	ا	Hamza/wasla alef variants → plain alef
  ة U+0629	ت	Ta marbuta → ta
  ۃ U+06C3	ہ	Small he → Urdu he
  ٮ U+066E	(removed)	Dotless ba — OCR artifact
  ٪ U+066A	%	Arabic percent → ASCII
  ZWJ / ZWNJ / LRM / RLM	(removed)	Invisible control characters

  Filtered out (whole sample dropped)

  • Any transcript containing digits (ASCII 0–9, Arabic-Indic ٠–٩, Extended ۰–۹)
  • Transcripts that are empty or fewer than 3 characters after cleaning
  • All non-Urdu / non-English script characters (Devanagari, Hebrew, CJK, etc.) are replaced with a space; samples that become empty are then dropped

  Space normalisation

  • Multiple consecutive spaces collapsed to a single space
  • Leading / trailing whitespace stripped

  Top 40 Characters (post-cleaning)

  Char	Codepoint	Unicode Name	Frequency
  ا	U+0627	ARABIC LETTER ALEF	1,764,365

| ی | U+06CC | ARABIC LETTER FARSI YEH | 1,656,310 | | ک | U+06A9 | ARABIC LETTER KEHEH | 996,578 | | و | U+0648 | ARABIC LETTER WAW | 921,501 | | ہ | U+06C1 | ARABIC LETTER HEH GOAL | 920,071 | | ر | U+0631 | ARABIC LETTER REH | 918,130 | | ے | U+06D2 | ARABIC LETTER YEH BARREE | 822,968 | | ن | U+0646 | ARABIC LETTER NOON | 669,620 | | م | U+0645 | ARABIC LETTER MEEM | 636,611 | | ت | U+062A | ARABIC LETTER TEH | 570,236 | | س | U+0633 | ARABIC LETTER SEEN | 555,358 | | ل | U+0644 | ARABIC LETTER LAM | 453,513 | | ں | U+06BA | ARABIC LETTER NOON GHUNNA | 427,813 | | ب | U+0628 | ARABIC LETTER BEH | 405,940 | | د | U+062F | ARABIC LETTER DAL | 307,807 | | ھ | U+06BE | ARABIC LETTER HEH DOACHASHMEE | 307,668 | | پ | U+067E | ARABIC LETTER PEH | 299,896 | | ج | U+062C | ARABIC LETTER JEEM | 257,909 | | گ | U+06AF | ARABIC LETTER GAF | 196,577 | | ٹ | U+0679 | ARABIC LETTER TTEH | 164,488 | | ئ | U+0626 | ARABIC LETTER YEH WITH HAMZA ABOVE | 147,302 | | چ | U+0686 | ARABIC LETTER TCHEH | 134,419 | | ف | U+0641 | ARABIC LETTER FEH | 123,204 | | ز | U+0632 | ARABIC LETTER ZAIN | 119,619 | | ع | U+0639 | ARABIC LETTER AIN | 109,658 | | ش | U+0634 | ARABIC LETTER SHEEN | 101,565 | | ق | U+0642 | ARABIC LETTER QAF | 90,511 | | ح | U+062D | ARABIC LETTER HAH | 76,356 | | ڈ | U+0688 | ARABIC LETTER DDAL | 72,531 | | آ | U+0622 | ARABIC LETTER ALEF WITH MADDA ABOVE | 68,159 | | ص | U+0635 | ARABIC LETTER SAD | 59,302 | | ۔ | U+06D4 | ARABIC FULL STOP | 57,814 | | خ | U+062E | ARABIC LETTER KHAH | 57,180 | | ڑ | U+0691 | ARABIC LETTER RREH | 51,705 | | ط | U+0637 | ARABIC LETTER TAH | 44,849 | | ض | U+0636 | ARABIC LETTER DAD | 24,586 | | ، | U+060C | ARABIC COMMA | 18,665 | | غ | U+063A | ARABIC LETTER GHAIN | 18,109 | | ظ | U+0638 | ARABIC LETTER ZAH | 16,061 |

## Schema

Each row contains the original columns from the source datasets plus the
cleaned `text` column. Key fields:

| Column | Type | Description |
|---|---|---|
| `text` | string | Normalised transcript |
| `duration` | float | Audio duration in seconds |

## License

Released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
Derived from source datasets — please also cite those accordingly.