Sarvam AI developed this synthetic benchmark in 2026 to evaluate context-aware Automatic Speech Recognition (ASR) within voice bot environments. The collection includes between 1,000 and 10,000 records covering the top 10 Indian languages, focusing on how conversation history and agent prompts influence transcription accuracy.
Use Cases
- Evaluating ASR accuracy using conversation history to improve transcription of user responses
- Fine-tuning models using agent prompts to provide linguistic context for speech-to-text tasks
- Benchmarking multilingual ASR performance across the 10 supported Indian languages
Strengths
- Covers 10 distinct Indian languages
- Provides conversation history and agent prompts for context-aware ASR testing
- Stored in optimized Parquet format for high-performance data loading
Limitations
- Synthetic generation may not capture the full range of real-world acoustic noise
- Small scale of 1,000 to 10,000 records limits its use for large-scale pre-training
- Scenario-specific focus on voice bots may not generalize to other speech domains
Provenance
- Source
- sarvamai
- Collection Method
- synthetic
- Freshness
- Last updated February 2026.
- Geography
- India