Hugging Face's logo

Datasets:
kaab4321
/
UrduTTS

Tasks:
Text-to-Speech
Automatic Speech Recognition
Languages:
Urdu
Size:
10K<n<100K
Tags:
urdu
speech
text-to-speech
tts
asr
low-resource
License:
Dataset card Data Studio Files
xet
 Community
1
UrduTTS / README.md
kaab4321's picture
kaab4321
Update README.md
a012374 verified
preview code
|
Raw History Blame Contribute Delete
11.4 kB
metadata 
language:
  - ur
license: cc-by-nc-4.0
pretty_name: 'UrduTTS: 92-Hour Urdu Speech Corpus with Three-Way Transcriptions'
size_categories:
  - 10K<n<100K
source_datasets:
  - original
task_categories:
  - text-to-speech
  - automatic-speech-recognition
annotations_creators:
  - machine-generated
language_creators:
  - found
multilinguality:
  - monolingual
tags:
  - urdu
  - speech
  - text-to-speech
  - tts
  - asr
  - low-resource
  - phonemes
  - ipa
  - romanization
  - south-asian-languages

UrduTTS

A Studio-Quality Urdu Speech Corpus with Urdu, Phonemized, and Romanized Transcriptions

91.9 hours · 57,873 utterances · 44.1 kHz · 3 aligned text representations

Paper GitHub License: CC BY-NC 4.0

Dataset Summary

UrduTTS is the largest openly available Urdu text-to-speech corpus with three aligned text representations for every utterance: native Urdu script, Phonemized (IPA) text, and Romanized (Latin) text.

Urdu is spoken by roughly 250 million people but is badly under-resourced in speech technology. Existing corpora are small, unreleased, or crowd-sourced with highly variable audio quality. UrduTTS provides 92 hours of clean, consistent read speech — and, critically, ships the phonemized layer that lets researchers move past the lossy Romanized transcriptions most Urdu TTS work has relied on.

Released with our LREC 2026 paper, which uses it to show that phonemized input outperforms romanized input on every metric measured, for both VITS and Glow-TTS.

Why three representations?

Urdu is phonetically rich: aspirated/unaspirated contrasts (/k/ vs /kʰ/), retroflex consonants (ٹ، ڈ، ڑ), nasalized vowels, and pharyngeal consonants from Arabic and Persian. Romanized Urdu — the input format most prior neural Urdu TTS systems used — encodes none of this reliably and has no orthographic standard: خواب ("dream") is written khawab, khwaab, or khwab depending on the writer.

Representation Field Example
Urdu (Perso-Arabic) transcription جس کی مثال بندگانِ نفس میں ناممکن ہے
Phonemized (IPA) phonemes ɟɪs ki mɪsˈaːl bˌəndəɡˈaːn nˈafs mˈẽ nˈaːmʊmkˌɪn hɛ
Romanized (Latin) roman JS KY MSAL BNDGAN NFS MYN NAMMKN HE

(English: "Whose example is impossible among the slaves of the self.")

The Romanized form drops the nasalized vowel in mˈẽ, the long vowels in mɪsˈaːl, and every stress mark. Quantifying that loss is what UrduTTS was built for. The clearest case: بڑی ("big") and بری ("bad") differ only by the retroflex ڑ, and both romanize to BRY — a contrast the model cannot learn because the input does not encode it.

Quickstart 

from datasets import load_dataset

ds = load_dataset("XKaab/UrduTTS")

sample = ds["train"][0]
sample["audio"]          # {'array': ndarray, 'sampling_rate': 44100, 'path': ...}
sample["transcription"]  # Urdu script
sample["phonemes"]       # IPA
sample["roman"]          # Romanized Latin

Stream it instead of downloading ~88 GB:

ds = load_dataset("XKaab/UrduTTS", split="train", streaming=True)
sample = next(iter(ds))

Resample to 22.05 kHz (the rate used for the paper's TTS training):

from datasets import Audio
ds = ds.cast_column("audio", Audio(sampling_rate=22050))

Play a sample in a notebook:

from IPython.display import Audio as Play
s = ds["train"][0]["audio"]
Play(s["array"], rate=s["sampling_rate"])

Dataset Structure

Fields

Field Type Description
audio Audio Mono 16-bit PCM WAV, 44.1 kHz
transcription string Native Urdu script (Perso-Arabic, Nastaliq)
phonemes string IPA phoneme sequence with stress marks, from eSpeak NG
roman string Deterministic Latin transliteration

Splits

Split Utterances Duration Share
train 40,511 64.21 h 70%
eval 11,575 18.41 h 20%
test 5,787 9.24 h 10%
Total 57,873 91.86 h

Utterance-level random split, seed=42, reproducible via data_pipeline/06_build_dataset.py.

Audio properties

Property Value
Format WAV, 16-bit PCM, mono
Sampling rate 44,100 Hz
Mean clip length 5.71 s
Clip length range 99.9% fall within 2–15 s (56 outliers, 0.19–41 s)
Mean words / chars per utterance 15.35 / 67.10
Speakers 4 (all male) — see limitations

How the corpus was built

Every stage is a runnable script in the GitHub repository.

  1. Sourcing — public-domain recordings (LibriVox, Internet Archive, Loyal Books), filtered for speaker consistency, pronunciation quality, and absence of background noise.
  2. Segmentation — split on silence with pydub: a pause longer than 400 ms below -15 dBFS marks a boundary; clips outside 2–15 s are rejected.
  3. Boundary synchronization — 100 ms is trimmed from the end of each clip and prepended to the next within the same recording, so hard cuts do not sever word onsets.
  4. Transcription — Urdu script via a Whisper-based pipeline (faster-whisper).
  5. Romanization — deterministic grapheme-to-Latin transliteration with a fixed table.
  6. Phonemization — IPA from eSpeak NG (ur), with stress marks preserved.

Benchmark results

VITS and Glow-TTS trained from scratch with Coqui TTS, once per text representation. Everything except the input representation is held constant.

Objective (500 samples per system)

System MCD Plain ↓ MCD DTW ↓ MCD DTW-SL ↓ SNR (dB) ↑ WER (%) ↓ CER (%) ↓
VITS-Ph 17.00 ± 2.71 7.97 ± 1.66 9.35 ± 2.19 61.86 12.3 5.3
VITS-Rm 17.53 ± 2.96 8.53 ± 1.74 9.98 ± 2.26 35.39 15.1 6.5
Glow-Ph 17.71 ± 2.01 6.74 ± 0.98 7.42 ± 1.26 10.40 12.5 5.8
Glow-Rm 18.83 ± 2.12 7.07 ± 1.07 7.86 ± 1.44 10.27 25.0 11.8

Subjective (MOS, 40 native listeners, 95% CI)

System MOS
Reference (ground truth) 4.79 ± 0.12
VITS-Phonemized 4.38 ± 0.24
VITS-Romanized 3.86 ± 0.36
GlowTTS-Phonemized 3.34 ± 0.28
GlowTTS-Romanized 2.64 ± 0.31

Phonemized input wins on every metric for both architectures: +9.6% MOS for VITS and +26.5% MOS for Glow-TTS. The clearest case is Glow-TTS intelligibility, where WER doubles under Romanization (25.0% vs 12.5%) — Romanized text cannot encode the retroflex and aspiration contrasts the model needs.

Intended Uses

Direct use — Urdu TTS and ASR, grapheme-to-phoneme and transliteration research, Urdu phonetics and prosody, and as a template for building resources in other under-resourced languages.

Out of scope

  • Commercial use of any kind — the license is CC BY-NC 4.0
  • Voice cloning or speaker impersonation of the corpus speakers
  • Generating deceptive synthetic speech attributed to a real person

Limitations

Please read before building on the corpus.

  • All four speakers are male, and speaker labels are not included in this release. Any voice trained on URDUTTS will be male, and per-speaker filtering or multi-speaker conditioning is not possible from the released fields alone. This is the corpus's most significant limitation.
  • Transcriptions are ASR-generated, not hand-verified. Urdu-script transcripts come from a Whisper-based pipeline and inherit its errors, which are non-trivial for Urdu. Because the Romanized and Phonemized layers are derived from those transcripts, an ASR error propagates to all three representations.
  • 9.8% of roman values contain residual Arabic diacritics (5,669 rows — mostly U+0650 kasra and U+0670 superscript alef), e.g. SFRِ HJ. The transliteration table does not map combining marks, so they pass through. Strip them with normalize_urdu() from data_pipeline/urdu_romanizer.py, or filter with re.sub(r'[^\x00-\x7F]', '', roman).
  • Read speech in a formal register. Sources are audiobook and literary recordings. Conversational Urdu, code-switching, and spontaneous speech are absent.
  • Romanization is a fixed convention, not a standard. It is deterministic and internally consistent, which is what makes the phonemized-vs-romanized comparison controlled. It deliberately reflects the lossiness of real Roman Urdu and is not a proposal for how Roman Urdu should be written.
  • Phonemization inherits eSpeak NG's Urdu limitations. Its Urdu front-end is rule-based and context-insensitive; short-vowel restoration in Perso-Arabic script is genuinely ambiguous (پر can be par or pur), so some phonemizations will be contextually wrong.

Relation to the paper

The paper's Table 2 reports 88.59 hours / 55,450 utterances at 22.05 kHz with a per-speaker breakdown. This release is a slightly larger build — 91.86 hours / 57,873 utterances — and stores audio at its native 44.1 kHz rather than the 22.05 kHz used for TTS training; resample on load with cast_column as shown above. Speaker labels used for the paper's per-speaker table are not part of the released fields.

Ethics

All source audio came from publicly available, public-domain recordings. The dataset contains no personally identifiable information and was anonymized before release. Synthetic Urdu speech built from this corpus should be disclosed as synthetic wherever it is deployed, and should not be used to produce speech attributed to real, identifiable individuals.

Citation 

@inproceedings{shahid-izharuddin-2026-urdutts,
    title     = {A Benchmark Dataset and Comparative Evaluation of Phonemized and
                 Romanized Urdu for Text-to-Speech},
    author    = {Shahid, M Kaab Bin and Izharuddin, Muhammed},
    booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation
                 Conference (LREC 2026)},
    year      = {2026},
    pages     = {10982--10993},
    publisher = {European Language Resources Association (ELRA)},
    url       = {https://github.com/KAABSHAHID/URDUTTS}
}

License

CC BY-NC 4.0 — share and adapt for non-commercial purposes with attribution. See creativecommons.org/licenses/by-nc/4.0. Code in the companion repository is MIT-licensed.

Code, pipeline, and training recipes: github.com/KAABSHAHID/URDUTTS