sarvamai/tts-general-benchmark

TTS General Benchmark

A multilingual Text-to-Speech (TTS) evaluation benchmark covering 11 Indian languages across multiple real-world use cases. The dataset is designed for systematic and repeatable evaluation of TTS systems under both high-quality and telephony bandwidth conditions.

Total prompts: 1,815 unique text samples Languages: 11 Evaluation tracks: High Quality + 8 kHz Telephony

This is an evaluation-only benchmark dataset intended for testing and comparison — not for model training.

Dataset Overview

The TTS General Benchmark provides diverse prompts that reflect practical deployment scenarios such as conversational agents, announcements, narration, support calls, and telephony bots. Prompts are curated to test clarity, robustness, pronunciation handling, and expressive capability.

Each prompt is labeled with:

• —language
• —usecase
• —eval_category (evaluation track)

The dataset contains two independently evaluated tracks with different prompt distributions.

Evaluation Categories

high_quality

Full-band prompts intended for studio / wideband TTS evaluation. These focus on naturalness, expressiveness, and content realism.

High Quality Use Cases

High-quality total: 1,265

8khz_telephony

Narrowband prompts designed for telephony and call-center evaluation (8 kHz playback target). These measure intelligibility, clarity, and robustness under bandwidth constraints.

Telephony Use Cases

Telephony total: 550

Supported Languages

Language coverage is shared across both evaluation tracks.

Dataset Structure

Each JSONL row contains:

{
  "text": "The text to be synthesized",
  "language": "hi",
  "usecase": "Conversational Bots",
  "eval_category": "high_quality"
}