README.md

metadata

license: apache-2.0
task_categories:
  - automatic-speech-recognition
language:
  - uz

Speech-to-Text Evaluation Dataset

Dataset Overview

This dataset is designed for evaluating Uzbek speech-to-text (STT) models on real-world conversational speech data. The audio samples were collected from various open Telegram groups, capturing natural voice messages in diverse acoustic conditions and speaking styles.

Key Statistics

• Total Samples: 745 audio files
• Total Duration: 1 hour 40 minutes (~100 minutes)
• Average Duration: ~8 seconds per sample
• Source: Voice messages from various open Telegram groups
• Transcriptions: Manually annotated

Dataset Structure

The dataset is saved as a datasets.Dataset object in Arrow format, containing the following fields:

• name: Name of audio file
• audio: Audio file data (dict with array, and sampling_rate)
• transcription: Ground truth text transcription (manually annotated)

Loading the Dataset

Installation

To use this dataset, you need to install the Hugging Face datasets library:

pip install datasets

Basic Loading

from datasets import load_dataset

# Load the dataset from the Arrow files
dataset = load_dataset("OvozifyLabs/asr_evaluate_set")

# View dataset information
print(dataset)
print(f"Number of samples: {len(dataset)}")

Data Characteristics

Audio Properties

• Source Domain: Conversational voice messages from Telegram
• Variability: Multiple speakers, diverse acoustic environments
• Recording Conditions: Real-world
• Language: Uzbek

Transcription Details

• Annotation Method: Manual transcription
• Quality: Human-verified ground truth labels
• Convention: punctuation removed, lowercased

Use Cases

This dataset is suitable for:

• Evaluating speech-to-text model performance on conversational speech
• Benchmarking ASR systems on real-world voice messages
• Testing model robustness to varied acoustic conditions
• Comparing different STT models