Vaani ASR Benchmark: Comprehensive Evaluation of Indian Language Speech Recognition

About the Vaani ASR Benchmark

The Vaani ASR Benchmark is a comprehensive evaluation framework designed to assess the performance of Automatic Speech Recognition (ASR) models across multiple Indian languages. This benchmark addresses the critical need for standardized evaluation of ASR systems in the linguistically diverse Indian subcontinent, where over 700 languages are spoken with 22 official languages recognized by the Constitution.

Why This Benchmark Matters

Addressing the Indian Language Gap: While significant progress has been made in ASR for high-resource languages like English and Mandarin, Indian languages have remained underrepresented in speech recognition research. The Vaani benchmark fills this critical gap by providing:

• Standardized Evaluation: Consistent metrics and methodology across different models and languages
• Diverse Linguistic Coverage: Support for major Indian languages including Hindi, Tamil, Telugu, Kannada, Bengali, and more
• Real-world Applicability: Evaluation datasets that reflect actual usage scenarios across India
• Research Acceleration: A common platform for researchers to compare and improve their ASR models

What We Evaluate

The benchmark evaluates ASR models across multiple dimensions:

🎯 Primary Metrics

• Word Error Rate (WER): Percentage of words incorrectly recognized (lower is better)
• Character Error Rate (CER): Percentage of characters incorrectly recognized (lower is better)

📊 Multiple Test Sets Our evaluation incorporates diverse, high-quality datasets:

1. FLEURS (Google): Multilingual speech corpus with 102 languages, providing ~10 hours per language with parallel sentences for robust cross-linguistic evaluation

2. Common Voice 12.0 (Mozilla): Community-contributed dataset with 26,119+ recorded hours across 104 languages, including rich demographic metadata (age, gender, accent)

3. IndicVoices (AI4Bharat): 12,000 hours of natural Indian speech covering 22 languages with diverse content:

   • Read speech (8%)
   • Extempore speech (76%)
   • Conversational speech (15%)
   • 22,563 speakers across 208 Indian districts

4. Gramvaani Hindi Dataset: Specialized Hindi ASR benchmark focusing on agriculture, healthcare, and general knowledge domains

5. MUCS 2021: Multilingual and code-switching dataset with ~600 hours across 7 Indian languages, including Hindi-English and Bengali-English code-switching

6. IndicTTS Database: 10,000+ utterances per language across 22 Indian languages with both native and English content

7. Kathbath (IndicSUPERB): 1,684 hours of labeled speech data across 12 Indian languages for comprehensive speech understanding evaluation

How We Evaluate

🔬 Rigorous Methodology Our evaluation follows a standardized protocol ensuring fair and accurate assessment:

Text Preprocessing Pipeline:

def clean(text):
    # Remove annotations and markup
    text = re.sub(r'{[^}]*}','',text)           # Remove {annotations}
    text = re.sub("[([].*?[)]]", "", text)      # Remove [brackets] and (parentheses)
    text = re.sub('<[^>]+>', '', text)          # Remove HTML/XML tags
    
    # Normalize punctuation
    text = text.replace("।", " ").replace("|", " ").replace("-", " ")\
        .replace(".", " ").replace(",", " ").replace("I", " ")\
        .replace('\n', ' ')
    
    # Normalize spacing
    text = re.sub(' +', ' ', text)
    return text.strip()

Error Rate Calculation:

• Uses industry-standard jiwer library for accurate WER/CER computation
• Identical preprocessing applied to both reference and hypothesis texts
• Results scaled to percentage (0-100) with 2-decimal precision
• Handles edge cases and missing data appropriately

Language Coverage

🗣️ Multilingual Support The benchmark currently supports major Indian languages with plans for expansion:

Currently Supported:

• Indo-Aryan: Hindi, Bengali, Marathi, Gujarati, Punjabi, Urdu, Assamese, Odia, Nepali
• Dravidian: Tamil, Telugu, Kannada, Malayalam
• Tibeto-Burman: Manipuri, Bodo
• Others: Sanskrit, Santhali

Planned Expansion:

• Additional regional languages and dialects
• Tribal and minority languages
• Code-switching scenarios (Hindi-English, Tamil-English, etc.)

Dataset Characteristics

📈 Comprehensive Coverage Our test datasets provide diverse evaluation scenarios:

Audio Quality Spectrum:

• Studio-quality recordings for controlled evaluation
• Real-world recordings capturing natural speech variations
• Telephonic and mobile recordings for practical applications

Speaker Diversity:

• Demographics: Balanced age, gender, and regional representation
• Accents: Multiple dialectal variations within languages
• Speaking Styles: Read speech, spontaneous speech, conversational audio

Content Variety:

• Domains: News, agriculture, healthcare, education, general knowledge
• Speech Types: Formal presentations, casual conversations, prompted responses
• Acoustic Conditions: Clean studio, noisy environments, multiple speakers

Performance Analysis

📊 Detailed Metrics

• AVG WER/CER: Simple average across all test datasets
• Language-specific Performance: Individual language breakdowns
• Dataset-specific Analysis: Performance variations across different test sets
• Statistical Significance: Confidence intervals and significance testing

🔍 Interactive Exploration

• Metric Selector: Switch between WER and CER views
• Language Filtering: Focus on specific languages or language families
• Dataset Comparison: Compare model performance across different test sets
• Trend Analysis: Track model improvements over time

Research Impact

🎯 Advancing Indian Language ASR The Vaani benchmark serves multiple stakeholders:

For Researchers:

• Standardized evaluation platform for model comparison
• Comprehensive datasets for training and testing
• Open-source evaluation code for reproducibility

For Industry:

• Performance benchmarks for commercial ASR systems
• Quality assurance metrics for product development
• Market readiness assessment for Indian language applications

For Society:

• Enabling voice interfaces in local languages
• Supporting digital inclusion across linguistic communities
• Preserving and promoting linguistic diversity through technology

Technical Implementation

🛠️ Robust Infrastructure

• Scalable Evaluation: Automated pipeline handling large-scale model evaluation
• Reproducible Results: Version-controlled datasets and evaluation scripts
• Quality Assurance: Multiple validation checkpoints and error detection
• Open Source: Full transparency in methodology and implementation

Future Roadmap

🚀 Continuous Enhancement

• Dataset Expansion: Adding more languages and domains
• Metric Refinement: Incorporating semantic and contextual evaluation measures
• Real-time Evaluation: Support for streaming ASR model assessment
• Community Integration: Enabling community contributions and model submissions

---

Citation

If you use this benchmark in your research, please cite:

@misc{vaani_asr_benchmark_2024,
  title={Vaani ASR Benchmark: Comprehensive Evaluation Framework for Indian Language Speech Recognition},
  author={Vaani Team},
  year={2024},
  url={https://vaani.iisc.ac.in}
}

For individual datasets used in the benchmark, please also cite the original sources as provided in our dataset documentation.