README.md

metadata

language:
  - multilingual
  - as
  - br
  - cy
  - et
  - eu
  - gl
  - hu
  - hy
  - ka
  - kk
  - lt
  - lv
  - mk
  - mt
  - oc
  - sk
  - sl
  - sw
  - ta
  - tk
  - tt
license: mit
task_categories:
  - automatic-speech-recognition

Whisper 3 Large Evaluation on Mozilla Common Voice 17 Rare Languages (Enhanced Metrics)

Dataset Description

This enhanced dataset contains comprehensive evaluation results of OpenAI's Whisper 3 Large model on rare languages from Mozilla Common Voice 17, with extensive additional metrics for thorough ASR evaluation.

Key Features

Enhanced Error Metrics:

• WER (Word Error Rate): Standard word-level error measurement
• CER (Character Error Rate): Character-level error measurement
• MER (Match Error Rate): Alternative error rate calculation
• WIL (Word Information Lost): Information loss measurement

Edit Distance Analysis:

• Word-level and character-level edit distances
• Normalized edit distance metrics
• Comprehensive distance analysis

Length and Structure Metrics:

• Word, character, and sentence counts
• Length ratios and differences
• Average word length analysis
• Sentence structure preservation

Script-Specific Analysis:

• Latin, Cyrillic, Armenian, Georgian, Tamil, Bengali character ratios
• Punctuation preservation analysis
• Script-specific performance metrics

Statistical Metrics:

• Jaccard similarity for vocabulary overlap
• Frequency correlation analysis
• Vocabulary union and overlap metrics
• Unique word analysis

Dataset Statistics

• Total samples: 111,507
• Languages: 21 rare languages
• Total metrics: 56 comprehensive evaluation metrics
• Scripts covered: Latin, Cyrillic, Armenian, Georgian, Tamil, Bengali

Language Coverage

Language	Code	Script	Sample Count
Assamese	as	Bengali	~551
Breton	br	Latin	~2,212
Welsh	cy	Latin	~5,379
Estonian	et	Latin	~2,653
Basque	eu	Latin	~13,630
Galician	gl	Latin	~9,990
Hungarian	hu	Latin	~11,435
Armenian	hy	Armenian	~4,281
Georgian	ka	Georgian	~12,618
Kazakh	kk	Cyrillic	~514
Lithuanian	lt	Latin	~4,753
Latvian	lv	Latin	~6,752
Macedonian	mk	Cyrillic	~1,097
Maltese	mt	Latin	~1,662
Occitan	oc	Latin	~254
Slovak	sk	Latin	~5,000
Slovenian	sl	Latin	~1,242
Swahili	sw	Latin	~12,253
Tamil	ta	Tamil	~12,074
Turkmen	tk	Latin	~546
Tatar	tt	Cyrillic	~4,964

Performance Highlights

Top Performing Languages (by WER):

1. Hungarian (hu): WER = 0.1822
2. Galician (gl): WER = 0.2027
3. Slovenian (sl): WER = 0.2205
4. Macedonian (mk): WER = 0.2762
5. Latvian (lv): WER = 0.3021

Usage

from datasets import load_dataset

# Load the enhanced dataset
dataset = load_dataset("norbertm/whisper-eval-rare-languages-csv")

# Access comprehensive metrics
print(dataset['train'][0])

Research Applications

This enhanced dataset enables:

1. Comprehensive ASR Evaluation: Multiple error metrics for thorough analysis
2. Script-Specific Analysis: Understanding performance across different writing systems
3. Statistical Analysis: Vocabulary and frequency correlation studies
4. Length Analysis: Understanding how text length affects recognition
5. Cross-Language Comparison: Detailed performance comparison across 21 languages

Citation

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

@dataset{whisper_eval_enhanced_2024,
  title={Whisper 3 Large Evaluation on Mozilla Common Voice 17 Rare Languages (Enhanced Metrics)},
  author={norbertm},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/norbertm/whisper-eval-rare-languages-csv}
}

License

This dataset is licensed under the MIT License.

---

This enhanced version includes 46 additional metrics beyond the original WER and CER, providing unprecedented depth for ASR evaluation research.