VoxParadox: Adversarial Speech QA Benchmark for Paralinguistic Understanding

An adversarial speech QA benchmark for evaluating paralinguistic understanding in Audio LLMs. The dataset was created by researchers from the University of Southern California and is hosted by IHP-Lab. It was last updated on May 28, 2026.

Use Cases

• Benchmarking Audio LLM robustness based on controlled linguistic-acoustic contradictions
• Training models to prioritize acoustic cues over contradictory transcripts
• Analyzing paralinguistic failure modes in speech understanding systems
• Mitigating model biases that over-rely on textual information

Strengths

• Examples are built around a controlled linguistic-acoustic contradiction, providing a clear test case
• The audio reliably conveys the correct paralinguistic attribute, ensuring a ground truth
• Hosted on Hugging Face, a major platform for AI datasets

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

IHP-Lab, University of Southern California

Collection Method

Likely constructed as a controlled benchmark with adversarial examples

Freshness

Last updated 2026-05-28 05:19:56; freshness should be verified