Dataset Viewer
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Parquet error: Scan size limit exceeded: attempted to read 363732221 bytes, limit is 300000000 bytes Make sure that 1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files
Error code:   TooBigContentError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Natural ASR Speech Samples

This set contains natural two-person conversations recorded in Hinglish, Tamil, and Bengali, paired with high-quality human-generated transcriptions. The recordings capture spontaneous, real-life dialogue including pauses, fillers, overlaps, and informal phrasing, making it ideal for building robust Automatic Speech Recognition (ASR) systems. Also we can support all indic languages and accents across India + Arabic + Korean + Vietnamese + Portuguese


Dataset Features

  • Natural, unscripted two-speaker conversations
  • Hinglish, Tamil, and Bengali multilingual coverage
  • Varied speaking speeds, tones, and regional accents
  • Clean and accurate human transcriptions
  • Includes pauses, interruptions, and conversational flow
  • Suitable for research and commercial ASR development with attribution

Intended Uses

✅ Direct Use

  • Training multilingual and code-mixed ASR models
  • Benchmarking conversational ASR performance
  • Hinglish language modeling
  • Accent-robust ASR system development
  • Dialogue understanding and speech-to-text tasks
  • Evaluation of spontaneous-speech ASR accuracy

❌ Out-of-Scope Use

  • Speaker or biometric identification
  • Psychological, emotion, or behavior profiling
  • Medical or clinical speech analysis
  • Commercial deployment without CC BY 4.0 credit
  • Real-time mission-critical ASR applications

Considerations and Limitations

  • ❗ Dataset size is limited (<1,000 samples) and may not include all dialects
  • 🎧 Contains fillers, hesitations, and overlapping conversation (true natural speech)
  • 🗣️ Accent diversity exists but is not fully representative of all regions
  • 🔄 Future versions will add more speakers, languages, and recording environments

License

CC BY 4.0 — Free to use, modify, distribute, and publish with attribution.


Contact

For dataset collaboration, contribution, or citation details, contact:

Downloads last month
42