You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

BabelSpeech: 50 Hours of Real-World Colloquial Malay ASR Speech Data

This dataset contains 50 hours of high-quality Malay colloquial ASR speech data, reflecting realistic code-switching between Malay and English, as commonly used in everyday communication in Malaysia.

Overview

  • Content: 50 hours of real-world colloquial Malay speech suitable for ASR fine-tuning and benchmarking.
  • Metadata: Stored in a separate JSON file, including audio path, duration, text, confidence score, SNR, and DNSMOS. More metadata fields may be added in future updates.
  • Domains covered: Technology, entertainment, travel, education, daily life, and others.
  • Data quality: Each batch was manually validated by native speakers. The transcription accuracy exceeds 98%, making it suitable for ASR model fine-tuning.
  • Sampling rate: 16 kHz

Data Selection and Quality Control Process

  1. Data sourcing — Collected from multiple high-quality open sources through international collaborations.
  2. Human screening — All samples are manually reviewed to remove noisy or low-quality segments.
  3. Automated preprocessing — Includes denoising, enhancement, and structured data cleaning.
  4. Audio quality evaluation — Automated tools evaluate SNR and DNSMOS to select the best-quality clips.
  5. Cross-model validation — Multi-model consistency checks are used to minimize transcription bias.
  6. Final human verification — Native speakers perform spot checks to ensure overall accuracy above 95%.

Metadata Format

{
  "filename": "youtube_ma_spk005_Rcl6dRZg-Os_202509_0013_00.wav",
  "relative_path": "wav/youtube_ma_spk005_Rcl6dRZg-Os_202509_0013_00.wav",
  "duration": 5.4,
  "confidence": 1.0,
  "text": " Ada dua lampu dekat sini. Ada dua lampu untuk suluh dia time dia streaming lah kot.",
  "snr": 54.93885803222656,
  "dnsmos": 4.4429564
}

Data Distribution

  • SNR Distribution

  • DNSMOS Distribution

  • Duration Distribution

  • Confidence Distribution

Contact

For inquiries or to request access to additional high-quality Malay speech data, please contact our team at: 📧 babelspeechai@gmail.com

Downloads last month
14