IHP-Lab/VoxParadox

Do Audio LLMs Listen or Read? Analyzing and Mitigating Paralinguistic Failures with VoxParadox

Jiacheng Pang^*, Ashutosh Chaubey^*, Mohammad Soleymani
University of Southern California
_{^* Equal contribution}

---

An adversarial speech QA benchmark for paralinguistic understanding in Audio LLMs. Each example is built around a controlled linguistic–acoustic contradiction: the transcript explicitly asserts an incorrect paralinguistic attribute, while the audio reliably conveys the correct one. Models that defer to transcript content are misled; models that listen are not.

2,000 MCQs across 10 paralinguistic tasks (200 each).

Quick start

from datasets import load_dataset
ds = load_dataset("IHP-Lab/VoxParadox", split="test")
print(ds[0])           # includes `audio` (decoded), `question`, `choice_a..d`,
                       # `answer_gt`, `adversarial_labels`, `task_name`, `id`

Tasks

y_true (audio) and y_adv (transcript) are disjoint by construction.

Task (task_name)	Acoustic attribute the model must recover
age_prediction	Speaker's age group
gender_prediction	Speaker's gender
emotion_recognition	Emotion conveyed by delivery (high-contrast pairs)
intonation_perception	Rising vs. falling intonation
speaker_identity_recognition	Which segment shares a speaker with a queried segment
total_speaker_counting	Number of distinct speakers
pitch_comparison	Ordering of three segments by pitch
volume_comparison	Ordering of three segments by loudness
speed_comparison	Ordering of three segments by speaking rate
vocal_range_comparison	Ordering of three segments by pitch range

File layout

.
├── metadata.jsonl    # one record per example (loaded by `datasets`)
├── voxparadox.json   # same content as JSON array (for direct inspection)
├── audio/            # 2,000 wav files
├── eval.py           # evaluation script (GT accuracy + ALA)
├── assets/           # logo
└── LICENSE           # USC Research License

Record schema

Field	Type	Description
id	string	{task_name}__{N}, with N running 0–199 within each task.
task_name	string	One of the 10 tasks above.
file_name / audio_path	string	Path to the audio clip, relative to this directory.
question	string	The MCQ question prompt.
choice_a / choice_b / choice_c / choice_d	string	The four answer options.
answer_gt	string	The acoustic ground-truth y_true (one of the four choices).
adversarial_labels	list[string]	The transcript-implied label(s) y_adv. Single-element for most tasks; 2 elements for the four *_comparison tasks.

Evaluation

Two complementary metrics:

• GT Accuracy — fraction matching answer_gt. Higher is better; reflects use of acoustic evidence.
• Adversarial-Label Agreement (ALA) — fraction matching any string in adversarial_labels. Higher ALA means more transcript-following.

Run eval.py on a JSONL of model predictions (one record per line, fields id and response):

python eval.py --predictions preds.jsonl

Example prediction record:

{"id": "age_prediction__0", "response": "(C) Elderly adult."}

The script parses A/B/C/D from the response (letter-first, then choice-text fallback), prints per-task and overall GT/ALA, and optionally writes a JSON report with --report report.json.

License

Released under the USC Research License (research and non-profit use only; commercial use requires a separate license). See LICENSE for the full text.

Audio was synthesized via commercial TTS engines (ElevenLabs, GPT-4o, Microsoft Azure); any commercial reuse of the audio is additionally subject to those vendors' terms of service.

Citation

@inproceedings{pang2026voxparadox,
  title     = {Do Audio LLMs Listen or Read? Analyzing and Mitigating Paralinguistic Failures with VoxParadox},
  author    = {Pang, Jiacheng and Chaubey, Ashutosh and Soleymani, Mohammad},
  booktitle = {Proceedings of the International Conference on Machine Learning (ICML)},
  year      = {2026}
}