Datasets:
metadata
pretty_name: TTS Robustness Benchmark
task_categories:
- text-to-speech
language:
- hi
- ta
- kn
- te
- ml
- en
- od
- gu
- mr
- bn
- pa
license: other
size_categories:
- 1K<n<10K
TTS Robustness Benchmark
Overview
The TTS Robustness Benchmark is a set of evaluation samples used to calculate domain-wise CER (Character Error Rate) scores across 7 critical stress-test categories. This benchmark ensures that each sentence-language pair appears exactly once, providing a clean and reliable metric for TTS robustness.
- Blog Post: Read the official announcement
Column Descriptions
| Column | Description |
|---|---|
text |
The original input sentence or phrase to be synthesized/evaluated. |
language |
The ISO-style language code (e.g., hi, od, ta). |
category |
The high-level functional domain (e.g., Numerics, STEM). |
Language Coverage
Total Samples: 959
| Language | Code | Sample Count |
|---|---|---|
| Hindi | hi |
103 |
| Tamil | ta |
97 |
| Kannada | kn |
94 |
| Telugu | te |
92 |
| Malayalam | ml |
85 |
| English | en |
84 |
| Odia | od |
81 |
| Gujarati | gu |
81 |
| Marathi | mr |
81 |
| Bengali | bn |
81 |
| Punjabi | pa |
80 |
Category (Domain) Distribution (7 Core Domains)
| Domain | Description | Count |
|---|---|---|
| Numerics | Phone numbers, currency, dates, time, complex decimals. | 267 |
| STEM | Chemical formulas, SI units, mathematical formulas. | 245 |
| Indian Named Entities | Brands, medicines, people, and place names. | 215 |
| Abbreviations | Acronyms and technical abbreviations. | 71 |
| Code-mixed | Sentences blending multiple languages (e.g., Hinglish). | 70 |
| Romanized | Indic languages written in Latin script. | 53 |
| URLs/Emails/Handles | Digital identifiers and web links. | 38 |