UrduSpeech is a high-quality multi-style speech corpus for Urdu and Kashmiri languages, designed for text-to-speech (TTS), automatic speech recognition (ASR), and expressive speech synthesis tasks. This dataset contains professionally recorded audio with diverse speaking styles, emotional expressions, and gender representation.

Dataset Composition

Languages: Urdu, Kashmiri
Total Samples: ~51.6K audio-text pairs
- Train: 46.5K samples
- Test: 5.17K samples
Dataset Size: 28.9 GB
Audio Quality: High-quality studio recordings
Gender Diversity: Male and Female speakers
Speaking Styles: Conversational (CONV), Fear (FEAR), Wikipedia (WIKI), Book narration (BOOK), and more

Source

This dataset is derived from AI4Bharat's Rasa Dataset, specifically the Urdu and Kashmiri language subsets.

Use Cases

This dataset is ideal for:

🔊 Text-to-Speech (TTS) - Train high-quality Urdu/Kashmiri speech synthesis models
🎤 Automatic Speech Recognition (ASR) - Develop robust speech-to-text systems
🎭 Expressive Speech Synthesis - Generate speech with different styles and emotions
👥 Multi-Speaker Models - Build gender-aware voice systems
📚 Domain-Specific TTS - Train models for conversational, narrative, or formal speech
🧠 Emotion Recognition - Develop systems to detect emotional expression in speech
🎯 Style Transfer - Learn to generate speech in different speaking styles

Data Fields

Field	Type	Description
`filename`	string	Unique identifier for the audio sample (e.g., URD_F_CONV_00045)
`text`	string	Urdu/Kashmiri transcription text
`language`	string	Language label (Urdu/Kashmiri)
`gender`	string	Speaker gender (Male/Female)
`style`	string	Speaking style (CONV, FEAR, WIKI, BOOK, etc.)
`duration`	string	Audio duration in seconds
`wav_path`	string	Original file path reference
`audio`	audio	Audio data object

Speaking Styles

The dataset includes multiple speaking styles to enable expressive TTS:

CONV (Conversational): Natural, everyday speech patterns
FEAR (Fearful): Emotional expression conveying fear or anxiety
WIKI (Wikipedia): Formal, encyclopedic narration style
BOOK (Book Narration): Literary, storytelling style
Additional styles may be present in the dataset

Data Examples

Example 1: Conversational Style

{
  'filename': 'URD_F_CONV_00045',
  'text': 'سچ تو یہ ہے، کہ ہم ہی بیوقوف تھے۔',
  'language': 'Urdu',
  'gender': 'Female',
  'style': 'CONV',
  'duration': '2.739',
  'wav_path': '/data/TTS/ttsteam/datasets/rasa_hf/data_20251203/Urdu/Female/wavs/URD_F_CONV_00045.wav',
  'audio': {...}
}

Example 2: Wikipedia Style

{
  'filename': 'URD_F_WIKI_01270',
  'text': 'احمد شاہ ابدالی نے کُل سات حملے کیے جن میں سے چوتھے اور پانچویں حملے میں دہلی پہنچا۔',
  'language': 'Urdu',
  'gender': 'Female',
  'style': 'WIKI',
  'duration': '6.525',
  'wav_path': '/data/TTS/ttsteam/datasets/rasa_hf/data_20251203/Urdu/Female/wavs/URD_F_WIKI_01270.wav',
  'audio': {...}
}

Data Splits

Split	Samples	Description
train	46,500	Training data for model development
test	5,170	Held-out test set for evaluation

Usage

Loading the Dataset

from datasets import load_dataset

# Load full dataset
dataset = load_dataset("humairawan/UrduSpeech")

# Load specific split
train_data = dataset['train']
test_data = dataset['test']

# Access a sample
sample = train_data[0]
print(f"Text: {sample['text']}")
print(f"Gender: {sample['gender']}")
print(f"Style: {sample['style']}")
print(f"Duration: {sample['duration']} seconds")

Filter by Language

from datasets import load_dataset

# Load dataset
dataset = load_dataset("humairawan/UrduSpeech", split="train")

# Filter for Urdu samples only
urdu_data = dataset.filter(lambda x: x['language'] == 'Urdu')

# Filter for Kashmiri samples only
kashmiri_data = dataset.filter(lambda x: x['language'] == 'Kashmiri')

print(f"Urdu samples: {len(urdu_data)}")
print(f"Kashmiri samples: {len(kashmiri_data)}")

Filter by Gender and Style

from datasets import load_dataset

dataset = load_dataset("humairawan/UrduSpeech", split="train")

# Get female conversational speech
female_conv = dataset.filter(
    lambda x: x['gender'] == 'Female' and x['style'] == 'CONV'
)

# Get male Wikipedia narration
male_wiki = dataset.filter(
    lambda x: x['gender'] == 'Male' and x['style'] == 'WIKI'
)

print(f"Female conversational samples: {len(female_conv)}")
print(f"Male Wikipedia samples: {len(male_wiki)}")

Filter by Duration

from datasets import load_dataset

dataset = load_dataset("humairawan/UrduSpeech", split="train")

# Get samples between 2-10 seconds
filtered = dataset.filter(
    lambda x: 2.0 <= float(x['duration']) <= 10.0
)

print(f"Filtered samples: {len(filtered)}")

Example: Fine-tuning TTS Models

from datasets import load_dataset
import torch
from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech

# Load model and processor
processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts")

# Load dataset
dataset = load_dataset("humairawan/UrduSpeech", split="train")

# Filter high-quality samples (2-10 seconds)
dataset = dataset.filter(lambda x: 2.0 <= float(x['duration']) <= 10.0)

# Preprocess function
def prepare_dataset(batch):
    audio = batch["audio"]
    batch["input_ids"] = processor(text=batch["text"], return_tensors="pt").input_ids
    batch["labels"] = processor(
        audio=audio["array"],
        sampling_rate=audio["sampling_rate"],
        return_tensors="pt"
    ).input_values
    return batch

# Process dataset
dataset = dataset.map(prepare_dataset)

Example: Style-Conditioned TTS Training

from datasets import load_dataset

# Load dataset
dataset = load_dataset("humairawan/UrduSpeech", split="train")

# Create style-specific subsets
styles = ['CONV', 'WIKI', 'BOOK', 'FEAR']
style_datasets = {
    style: dataset.filter(lambda x: x['style'] == style)
    for style in styles
}

# Train separate models or use style embeddings
for style, style_data in style_datasets.items():
    print(f"{style}: {len(style_data)} samples")

Example: ASR Training with Whisper

from datasets import load_dataset
from transformers import WhisperProcessor, WhisperForConditionalGeneration

# Load processor and model
processor = WhisperProcessor.from_pretrained("openai/whisper-small")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")

# Load dataset
dataset = load_dataset("humairawan/UrduSpeech", split="train")

# Preprocessing function
def prepare_asr_dataset(batch):
    audio = batch["audio"]
    # Process audio
    batch["input_features"] = processor(
        audio["array"],
        sampling_rate=audio["sampling_rate"],
        return_tensors="pt"
    ).input_features[0]
    # Process text
    batch["labels"] = processor.tokenizer(batch["text"]).input_ids
    return batch

# Apply preprocessing
dataset = dataset.map(prepare_asr_dataset, remove_columns=["audio"])

Dataset Statistics

Total Duration: Extensive audio coverage across both languages
Average Duration: ~5-7 seconds per sample
Gender Distribution: Balanced male and female representation
Style Diversity: Multiple speaking styles for expressive synthesis
Language Coverage: Comprehensive Urdu and Kashmiri lexicon
Audio Quality: High-fidelity studio recordings

Naming Convention

Audio filenames follow the pattern: {LANGUAGE}_{GENDER}_{STYLE}_{ID}

Example: URD_F_CONV_00045

URD = Urdu language
F = Female speaker
CONV = Conversational style
00045 = Sample ID

Licensing & Attribution

This dataset is released under the CC-BY-4.0 license.

Source: Derived from AI4Bharat's Rasa Dataset

Citation:

@dataset{urduspeech2025,
  title        = {UrduSpeech: Multi-Style Urdu and Kashmiri Speech Corpus},
  author       = {Humair Munir},
  author       = {humair025},
  year         = {2025},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/humairawan/UrduSpeech},
  note         = {Derived from AI4Bharat Rasa Dataset}
}

Please also cite the original Rasa dataset:

@misc{ai4bharat2024rasa,
  title        = {Rasa: Building Expressive Speech Synthesis Systems for Indian Languages in Low-Resource Settings},
  author       = {AI4Bharat},
  year         = {2024},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/ai4bharat/Rasa}
}

Ethical Considerations

This dataset is intended for research and development purposes
Users should ensure compliance with privacy regulations when deploying models
The dataset includes emotional expressions (fear) - use responsibly
Consider cultural context when deploying applications built with this data
Ensure appropriate content warnings for applications using emotional speech

Limitations

Speaker diversity is limited to the original Rasa dataset speakers
Not all speaking styles may be evenly distributed
Kashmiri language samples may be fewer than Urdu
Domain coverage is limited to the source material styles
Some emotional expressions may not represent all cultural contexts

Technical Specifications

Audio Format: WAV format
Recommended Sampling Rate: 16kHz or 22kHz for TTS applications
Text Encoding: UTF-8 (Urdu script)
File Format: Parquet
Total Size: 28.9 GB

Recommended Preprocessing

# Recommended filtering for TTS training
filtered = dataset.filter(
    lambda x: 2.0 <= float(x['duration']) <= 10.0
)

# For ASR, you might want shorter samples
asr_filtered = dataset.filter(
    lambda x: 1.0 <= float(x['duration']) <= 15.0
)

# For style-specific models
style_filtered = dataset.filter(
    lambda x: x['style'] in ['CONV', 'WIKI', 'BOOK']
)

Related Datasets

UrduMegaSpeech-1M

For large-scale Urdu speech recognition and translation tasks, see our companion dataset:

🔗 UrduMegaSpeech-1M - A large-scale Urdu-English parallel speech corpus with 1M+ samples

When to use which dataset:

Feature	UrduSpeech (This Dataset)	UrduMegaSpeech-1M
Size	51.6K samples	1M+ samples
Primary Use	TTS, Expressive Speech	ASR, Translation
Languages	Urdu, Kashmiri	Urdu-English parallel
Speaking Styles	Multiple (CONV, WIKI, BOOK, FEAR)	General domain
Quality	High-quality studio recordings	Variable quality with scores
Best For	Voice synthesis, style transfer	Large-scale ASR, translation

Recommendation:

Use UrduSpeech for text-to-speech, expressive synthesis, and style-conditioned models
Use UrduMegaSpeech-1M for robust ASR training, speech translation, and when you need large-scale data

Acknowledgments

This dataset is curated from AI4Bharat's Rasa project, which aims to build expressive speech synthesis systems for Indian languages. We acknowledge their contribution to low-resource language technology.

Dataset Curated By: Humair Munir
Original Source: AI4Bharat Rasa Dataset
Last Updated: December 2025
Version: 1.0

Downloads last month: 42