PishangShedappp/Malaysia-Whisper-Berita

Malaysia Whisper Berita Dataset

Dataset Description

This is a custom audio dataset designed specifically for fine-tuning Automatic Speech Recognition (ASR) models, such as OpenAI's Whisper, on the Malay language (Bahasa Melayu).

The audio consists of high-quality Malaysian news broadcasts (Berita), which provide excellent examples of formal, standard Malay (Bahasa Baku) and domain-specific vocabulary (e.g., politics, economy, current events).

• Language: Malay (ms-MY)
• Task: Automatic Speech Recognition (ASR) / Speech-to-Text (STT)
• Format: Audio clips (16kHz) with corresponding text transcriptions.

Purpose

Standard Whisper models are often trained on casual or general internet data. This dataset was created to fine-tune models to accurately transcribe formal Malaysian news broadcasts with high precision, making the resulting models ideal for deployment in local transcription tools or utilizing lightweight frameworks like faster-whisper.

Dataset Structure

The dataset is structured using Hugging Face's audiofolder format. It contains:

• audio: Short audio segments (typically 0-30 seconds).
• sentence: The text transcription of the audio segment.

All audio has been pre-processed and resampled to 16kHz to meet the input requirements for Whisper training.

Usage

You can load this dataset directly using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("PishangShedappp/Malaysia-Whisper-Berita")
print(dataset)