README.md

VOA PT-BR Raw Audio Dataset

Dataset Description

This dataset contains raw audio chunks processed directly with Whisper (without VAD preprocessing). Each chunk has an audio file (.wav) and associated metadata including:

• Raw transcription (automated using Whisper)
• Duration
• Start and end times
• Language (pt-br)

Usage

This dataset is intended for:

1. Manual review and correction of transcriptions
2. Evaluation of direct transcription quality
3. Fine-tuning of ASR models for Portuguese
4. Future comparison with VAD preprocessed audio

Dataset Structure

• audio/: Contains WAV files of audio chunks
• metadata.json: Contains complete metadata for all chunks
• metadata.csv: CSV version of the complete metadata
• simple_dataset.csv: Simplified version with just audio and transcription columns
• dataset_dict.json: Dataset in HuggingFace compatible format

Loading the Dataset

You can load this dataset using the HuggingFace Datasets library:

from datasets import load_dataset

# Load from the Hub
dataset = load_dataset("voa-engines/audio-review-raw")

# Access the first example
example = dataset["train"][0]
print(f"Audio file: {example['audio']}")
print(f"Transcription: {example['transcription']}")

Processing Information

The audio files were processed using the following approach:

• Direct Whisper transcription without VAD preprocessing
• Long audio files (>30s) were chunked into 30-second segments
• Original audio file structure maintained

Created on: 2025-04-04