SteerBench

SteerBench is a benchmark for spoken steerability. It tests whether a speech-to-speech assistant follows an explicit spoken request to change tone, persona, style/accent or speed/length while doing the requested task.

• 📄 Paper: SteerDuplex: Steerable Duplex Speech Dialogue Models (arXiv:2609.12623)
• 💻 Code / inference / judge: github.com/Utkarsh4430/SteerDuplex (steerbench/)

The whole benchmark is in a single self-contained parquet: every item's text, its rubrics, and the user turn + reference clip (audio embedded, playable in the viewer). The system prompt and voice prompt used to run the benchmark are alongside it.

Usage

This dataset is gated: click to accept the terms above, then authenticate with huggingface-cli login.

from datasets import load_dataset, Audio
ds = load_dataset("utkarsh4430/SteerBench", split="test")   # 390 rows: text, rubrics, embedded audio

# Text and rubrics work directly. Decoding the embedded audio needs torchcodec (pip install torchcodec),
# or read the raw WAV bytes without it:
ds_bytes = ds.cast_column("user_audio", Audio(decode=False)).cast_column("reference_audio", Audio(decode=False))
wav = ds_bytes[0]["user_audio"]["bytes"]

Each row has: user_audio, reference_audio (embedded audio), user_utterance, transcript, tts_instruct, steering_axis / sub_axis, and the rubrics (audio_rubrics, text_content_rubrics, text_constraint_rubrics).

path	contents
data/test-*.parquet	the whole benchmark: item text, rubrics, and both audio clips embedded (viewer-browsable)
system_prompt.txt	the examinee system prompt used to run the benchmark
voice_prompt/	the voice prompt WAV (for voice-conditioned models)

• Licence. Benchmark data: CC BY-NC-SA 4.0 (non-commercial, share-alike); see LICENSE. The code is MIT.
• Code (inference, judge, scorer) is in the GitHub repository (Utkarsh4430/SteerDuplex, steerbench/); its scripts/fetch_dataset.py downloads this dataset into data/.

Citation

If you use SteerBench, please cite:

@misc{tyagi2026steerduplexsteerableduplexspeech,
      title={SteerDuplex: Steerable Duplex Speech Dialogue Models},
      author={Utkarsh Tyagi and Ramaneswaran Selvakumar and Advait Gosai and Sonal Kumar and Nikhil Barhate and Isabell Sagar and Steven Li and Miheer Bavare and Daniel Quigley and Fabiola Tapia Carrillo and Jose M Patron E and Diego Macías Gutiérrez and Paul Song and Ramani Duraiswami and Dinesh Manocha and Yunzhong He},
      year={2026},
      eprint={2609.12623},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2609.12623},
}