Aqiba/memoni_20sec_segments

Memoni Audio Dataset - 20 Second Segments

Dataset Description

This dataset contains 3,058 audio segments derived from the original Aqiba/memoni_clean_audio dataset. Each audio segment is exactly 20 seconds long, making it compatible with the Sarvam AI REST API (which has a 30-second limit per request).

Dataset Summary

• Source: YouTube videos (Memoni/Gujarati/Urdu content)
• Original files: 371 audio files (avg. 82 seconds each)
• Segments created: 3,058 (each exactly 20 seconds)
• Audio format: 16kHz mono
• Purpose: Speech recognition, transcription, and audio processing

Languages

• Memoni (primary)
• Gujarati
• Urdu
• Some Arabic (Islamic prayers/phrases)

Dataset Structure

Data Fields

Field	Type	Description
audio	Audio	Audio array with 16kHz sampling rate
video_id	string	Source YouTube video ID
original_url	string	Full YouTube URL
segment_id	int32	Segment index within original file (0, 1, 2, ...)
global_segment_id	int32	Unique identifier across all segments
segment_start_sec	float32	Start time in original file (seconds)
segment_end_sec	float32	End time in original file (seconds)
duration_seconds	float32	Always 20.0 (padded if needed)
original_duration_seconds	float32	Original file duration before segmentation

Data Splits

Split	Size
Train	3,058 segments

Usage Example

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("Aqiba/memoni_20sec_segments", split="train")

# Access a sample
sample = dataset[0]
print(f"Video ID: {sample['video_id']}")
print(f"Segment ID: {sample['segment_id']}")
print(f"Duration: {sample['duration_seconds']} seconds")

# Audio is available as numpy array
audio_array = sample['audio']['array']
sampling_rate = sample['audio']['sampling_rate']

Dataset Creation

This dataset was created by:

1. Loading the original Aqiba/memoni_clean_audio dataset (371 files, variable duration)
2. Splitting each audio file into 20-second segments
3. Padding the last segment with silence if shorter than 20 seconds
4. Preserving all original metadata (video_id, original_url, etc.)

Segmentation Statistics

• Average segments per original file: ~8.2
• Total segments: 3,058
• Total audio hours: ~17 hours (3,058 × 20 seconds)

Contact

For questions or issues, please open an issue on the Hugging Face dataset repository.