SpeechArenaBench: Multilingual TTS Preference Data Across 10 Indian Languages

SpeechArenaBench is a large-scale human-preference dataset for evaluating multilingual Text-to-Speech systems. It contains benchmark sentences, generated audio, and crowd-sourced pairwise preference judgments collected from native raters across 10 Indian languages. The dataset was created by ai4bharat and accompanies a paper accepted to Interspeech 2026.

Use Cases

• Benchmarking TTS system performance based on human preference judgments mentioned in the description
• Training or fine-tuning TTS models based on crowd-sourced pairwise preference data
• Analyzing cross-lingual speech synthesis quality based on data from 10 Indian languages
• Studying human preferences in speech synthesis based on the collected native rater judgments

Strengths

• Covers 10 Indian languages, providing multilingual coverage
• Includes human-preference judgments collected from native raters
• Accompanies a peer-reviewed paper accepted to Interspeech 2026

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment
• Description metadata is limited; actual data quality requires manual inspection after download

Provenance

Source

ai4bharat

Collection Method

Crowd-sourced pairwise preference judgments collected from native raters

Freshness

Last updated 2026-06-18 10:26:21; freshness should be verified

Geography

India (10 Indian languages)