benkum/contextual_asr_benchmark

Synthetic Contextual ASR Benchmark (Indic)

Dataset Summary

This dataset is a Synthetic Contextual Automatic Speech Recognition (ASR) benchmark designed to evaluate and improve speech recognition systems in voice bot scenarios. It focuses on context-aware transcription, where the ASR model can leverage conversation history and agent prompts to better transcribe user responses.

The dataset covers the top 10 Indian languages, providing a diverse linguistic landscape for testing voice AI capabilities in real-world conversational settings.

Supported Languages

The dataset includes samples for the following 10 major Indian languages:

1. 1.Hindi (hi)
2. 2.Bengali (bn)
3. 3.Marathi (mr)
4. 4.Telugu (te)
5. 5.Tamil (ta)
6. 6.Gujarati (gu)
7. 7.Kannada (kn)
8. 8.Malayalam (ml)
9. 9.Odia (or)
10. 10.Punjabi (pa)

Dataset Structure

Data Instances

Each data instance represents a single turn in a voice bot interaction. The context field provides the necessary background (bot persona, history, and the immediate question) to help the model predict the text (transcription) from the audio.

Data Fields

• —`audio`: The audio file or data of the user's spoken response.
• —`text`: The ground truth transcription of the user's spoken response.
• —`language`: The language of the audio.
• —`context`: A text string containing the input scenario information, comprising:
• —Bot Description: The persona of the bot (e.g., "Banking Assistant").
• —Previous Conversation History: Previous turns in the dialogue.
• —Question asked by the bot: The specific query prompting the user's response.

Use Cases

This dataset is specifically designed for:

• —Contextual Biasing: Training ASR models to boost probabilities for expected words (e.g., numbers, dates, entities) based on the context.
• —Intent Recognition: Evaluating if the transcription captures the user's intent correctly in noisy scenarios.
• —Dialog State Tracking: Testing end-to-end spoken language understanding (SLU) systems.

Dataset Creation

• —Source: Synthetic generation.
• —Methodology: Voice bot scenarios were simulated to cover various domains (Banking, E-commerce, Healthcare). User responses were synthesized or recorded to match the specific prompt found in the context.

Citation

If you use this dataset, please cite it as follows:

@dataset{sarvam/contextual_asr_benchmark,
  author = {SarvamAi},
  title = {Synthetic Contextual ASR Benchmark: Top 10 Indian Languages},
  year = {2025},
  publisher = {Hugging Face}
}