Javed, Tahir, Janki Nawale, Sakshi Joshi, et al. 2024. “LAHAJA: A Robust Multi-Accent Benchmark for Evaluating Hindi ASR Systems.” arXiv:2408.11440. Preprint, arXiv, August 21. https://doi.org/10.48550/arXiv.2408.11440.

Notes

• 12.5 hrs Hindi audio
  • 132 speakers, 83 districts
• Data type
  • Read speech
    • 1K sentences from Wikipedia articles covering 13 domains
  • Digital interactions with voice assistants
    • in-home assistants for everyday tasks
    • digital payment services covering multiple intents
    • online grocery shopping apps covering multiple intents
    • online government services covering multiple intents
  • Extempore conversations
    • 2.5K questions from 21 domains such as tourism, government etc., and 28 topics of interest such as reading, painting etc
  • Named entities
    • users to speak any 5 numbers , any 5 dates, any 5 person names, names of any 5 Indian cities, any 5 Indian states, any 5 Indian districts, any 5 countries, and any 5 international cities

In-text annotations