README.md

Uzbek Speech Recognition (ASR) Dataset

This is an Uzbek language speech recognition dataset created for training ASR models.

Dataset Structure

• train: Training data (80% of conversations)
• validation: Validation data (10% of conversations)
• test: Test data (10% of conversations)

Audio Format

• Sample Rate: 16 kHz
• Channels: Mono
• Bit Depth: 16-bit PCM
• Format: WAV

Dataset Fields

• audio: Audio file (Audio feature, directly playable)
• text: Normalized text transcription in Uzbek (Latin script)
• duration: Duration of audio in seconds
• language: Language code (uzn_Latn for Uzbek Latin)
• id: Unique identifier for the segment
• audio_id: Identifier for the conversation/recording
• segment_index: Index of the segment within the conversation

Dataset Statistics

• Total conversations: 8
• Total segments: 48
• Total duration: 0.08 hours
• Train: 38 segments (0.06 hours)
• Validation: 0 segments (0.0 hours)
• Test: 10 segments (0.02 hours)

Important Notes

• Conversations are NOT split across train/validation/test sets
• Each conversation (audio_id) belongs entirely to one split
• This ensures proper model evaluation without data leakage
• Dataset is updated daily with new recordings (append-only)
• Once uploaded, segments are marked in database to prevent duplicates

Usage

Load with datasets library:

from datasets import load_dataset

# Load entire dataset
dataset = load_dataset('admin-euphoria/ASR_Dataset')

# Load specific split
train_dataset = load_dataset('admin-euphoria/ASR_Dataset', split='train')

# Access audio and text
for example in train_dataset:
    audio_array = example['audio']['array']
    sampling_rate = example['audio']['sampling_rate']
    text = example['text']
    print(f"Text: {text}, Duration: {example['duration']}s")

Update Strategy

• Daily building: First script collects data from database daily
• Incremental upload: Second script uploads only new segments to HF
• Database tracking: ProcessedDataset table tracks uploaded segment_ids
• No duplicates: Segment_id uniqueness prevents re-uploading
• Append-only: Data is only added, never removed or modified

Dataset Information

• Last Updated: 2025-12-06T16:15:27.710995
• Language: Uzbek (Latin script)
• Update Strategy: Incremental (append-only)
• License: [MIT: https://opensource.org/licenses/MIT]

Contact

[kkadyr039@gmail.com]