NisargBhavsar25/koel-benchmark

Koel Benchmark Suite

Dataset Description

The Koel Benchmark Suite is a comprehensive set of evaluation datasets designed to rigorously test the real-world performance of Text-to-Speech (TTS) models for major Indian languages. The suite focuses on challenges unique to the Indian context, such as code-switching, domain-specific terminology, proper nouns, and complex numeric formats.

This dataset was created to help developers and researchers build more natural, accurate, and culturally resonant voice systems for India. It addresses the gap left by generic benchmarks that fail to capture the nuances of everyday Indian speech.

For a detailed overview of the generation process, statistical analysis, and the motivation behind this project, please read our official announcement article: [LINK TO YOUR BLOG POST HERE]

Dataset Structure

The Koel Benchmark Suite is organized into four distinct splits, each targeting a specific TTS capability:

Data Splits

• koel_g (Generic): A foundational benchmark for evaluating a model's ability to handle natural, everyday code-switched conversations across various industries like BFSI, Healthcare, and Retail.
• koel_i (Industry): Tests a model's performance on specialized and often tricky vocabulary, including industry-specific jargon and prominent Indian company and brand names (e.g., "Tata Motors," "Paytm").
• koel_n (Numeric): Rigorously tests the synthesis of diverse numeric and alphanumeric formats critical for transactional and logistical applications, such as phone numbers, order IDs, PIN codes, and monetary amounts.
• koel_l (Locality): A special challenge dataset to evaluate the pronunciation of Indian place names. In this split, locality names are transliterated into the native script to test the model's phonetic understanding.

How to Use

The dataset can be easily loaded using the datasets library. You can load the entire collection or access a specific benchmark split by its name.

Load a Specific Benchmark

from datasets import load_dataset

# Load the generic conversational benchmark

koel_g_dataset = load_dataset("NisargBhavsar25/koel-benchmark", split='koel_g')

# Load the industry-specific benchmark

koel_i_dataset = load_dataset("NisargBhavsar25/koel-benchmark", split='koel_i')

# Example: Access the first sample from the numeric benchmark

sample = koel_n_dataset
print(sample['generated_sentence'])

Iterate Through All Benchmarks

from datasets import load_dataset

# Load all splits into a DatasetDict

all_koel_benchmarks = load_dataset("NisargBhavsar25/koel-benchmark")

print(all_koel_benchmarks.keys())

# dict_keys(['koel_g', 'koel_i', 'koel_n', 'koel_l'])

# Filter the locality benchmark for the Tamil language

tamil_locality_data = all_koel_benchmarks['koel_l'].filter(
lambda example: example['language'] == 'tamil'
)

Data Fields

All splits share a common set of columns, with specialized columns added for the industry, numeric, and locality benchmarks.

Field Name	Data Type	Description	Applicable Splits
language	string	The target language of the sentence (e.g., 'hindi', 'tamil').	All
generated_sentence	string	The final, generated sentence for TTS evaluation.	All
benchmark_type	string	The type of benchmark this sample belongs to.	All
emotion	string	The target emotional tone for the sentence (e.g., 'empathy', 'urgent').	All
industry	string	The business domain context (e.g., 'BFSI', 'Retail').	All
sentence_char_length	int64	The total number of characters in the generated sentence.	All
actual_word_count	int64	The total number of words in the generated sentence.	All
gender	string	The target gender of the speaker persona (always 'female').	All
scenario_type	string	The conversational context (e.g., 'sales', 'support').	All
use_complex_words	bool	A flag indicating if the prompt requested complex vocabulary.	All
generation_id	string	A unique identifier for the generation task.	All
company_name	string	The specific company or brand name included in the sentence.	koel_i only
number_value	string	The raw numeric or alphanumeric value.	koel_n only
number_type	string	The category of the number (e.g., 'phone_number', 'pincode', 'amount').	koel_n only
locality_name	string	The specific Indian locality or place name.	koel_l only
city	string	The city associated with the locality name.	koel_l only

Supported Languages

The benchmarks cover the following languages, often with English words mixed in to simulate code-switching:

• Hindi (hi)
• Tamil (ta)
• Telugu (te)
• Kannada (kn)
• Malayalam (ml)
• Marathi (mr)

Citation

If you use the Koel Benchmark Suite in your research or products, please cite it as follows:

@dataset{bhavsar_2025_koel,
author    = {Nisarg Bhavsar},
title     = {Koel: An Open-Source Benchmark Suite for Real-World Indian Language TTS},
month     = {July},
year      = {2025},
publisher = {Hugging Face},
version   = {1.0.0},
url       = {https://huggingface.co/datasets/NisargBhavsar25/koel-benchmark}
}