abidanoaman/urdu-asr-multitask-dataset

Urdu Multi-Task ASR Dataset

Preprocessed Urdu dataset for multi-task learning: ASR + Emotion + Gender

Dataset Details

• Language: Urdu (ur)
• Tasks: ASR (all samples), Emotion (subset), Gender (subset)
• Audio: Raw 16kHz mono WAV (ready for wav2vec2 fine-tuning)
• Total Samples: 7,155

Splits

Split	Samples	Emotion %	Gender %
Train	5,724	41.9%	48.8%
Validation	715	41.8%	48.8%
Test	716	41.9%	48.9%

Features

• audio: Raw audio array (16kHz mono)
• text: Urdu transcription
• emotion_label: Emotion ID (0-6) or -1 if missing
• gender_label: Gender ID (0-1) or -1 if missing
• has_asr: Task mask for ASR (always 1)
• has_emotion: Task mask for emotion (0 or 1)
• has_gender: Task mask for gender (0 or 1)
• source: Data source (common_voice or emotion)
• duration: Audio duration in seconds

Label Vocabularies

Emotion Labels

{ "0": "خوف", "1": "غصہ", "2": "خوشی", "3": "بوریت", "4": "افسردگی", "5": "ناگواری", "6": "غیر جانبدار" }

Gender Labels

{ "0": "مرد", "1": "عورت" }

Task Masks

Task masks enable proper multi-task learning:

• has_asr=1: Sample has ASR transcription (all samples)
• has_emotion=1: Sample has emotion label (emotion dataset only)
• has_gender=1: Sample has gender label (emotion dataset only)

Missing labels are encoded as -1, NOT fake labels!

Usage

from datasets import load_dataset

dataset = load_dataset("abidanoaman/urdu-asr-multitask-dataset")

# Check task masks
train_data = dataset['train']
print(f"ASR samples: {sum(train_data['has_asr'])}")
print(f"Emotion samples: {sum(train_data['has_emotion'])}")
print(f"Gender samples: {sum(train_data['has_gender'])}")

Training Strategy

1. Stage 1 (ASR-only): Train on all samples with ASR task
2. Stage 2 (Multi-task): Train on emotion/gender subset with all tasks

Use task masks to skip loss computation for missing labels!

Preprocessing

• Mono conversion
• 16kHz resampling
• Peak normalization
• Silence trimming (top_db=20)
• Duration: 0.5-20.0 seconds
• Stratified splits (balanced task distribution)

Citation

If you use this dataset, please cite the source datasets:

• Common Voice Urdu
• Urdu Emotion Voice Dataset