slprl/common-sense-facts-audio

Common-Sense Facts Audio Dataset

A spoken fact-completion dataset for evaluating whether Speech Language Models can retrieve common-sense and factual knowledge from speech.

Each example contains three paired versions:

1. prompt: an incomplete factual prompt, e.g. "the capital of France is"
2. fact: the correct full sentence, e.g. "the capital of France is Paris"
3. counterfactual: an incorrect matched sentence from the same category, e.g. "the capital of France is Rome"

The dataset covers categories such as capitals, colors, arithmetic, opposites, family relations, object functions, and common-sense facts.

It is designed for likelihood-based evaluation: a Speech LM succeeds when it assigns higher likelihood to the correct fact than to the matched counterfactual. The dataset also includes prompt-level audio and word-level timing information, making it useful for interpretability analyses of how speech-derived representations encode lexical and factual information.

Introduced in:

Interleaved Speech Language Models Latently Work In Text

📃 Paper | 🌐 Project Page

Usage

from datasets import load_dataset

ds = load_dataset("slprl/common-sense-facts-audio", split="data")

# Access a single example
row = ds[0]
print(row["id"])            # "colors_001"
print(row["category"])      # "Colors"
print(row["prompt_text"])   # "The color of grass is"
print(row["answer"])        # "green"

# Audio is a dict with array and sampling_rate
audio = row["prompt_audio"]   # {"array": np.ndarray, "sampling_rate": int}

# Filter by category
capitals = ds.filter(lambda x: x["category"] == "Capital cities")

---

Columns

Column	Type	Description
id	string	Clean example ID, e.g. capital_cities_001
category	string	Semantic category
prompt_text	string	Incomplete factual prompt
fact_text	string	Correct full sentence
counterfactual_text	string	Incorrect full sentence
answer	string	Correct answer word/phrase
counterfactual_answer	string	Incorrect answer word/phrase
prompt_audio	Audio	WAV of the prompt
fact_audio	Audio	WAV of the correct full sentence
counterfactual_audio	Audio	WAV of the incorrect full sentence
prompt_words	list[str]	Word tokens of the prompt
prompt_word_starts	list[float]	Word start times in seconds
prompt_word_ends	list[float]	Word end times in seconds (-1.0 = last word; use audio duration)

---

Examples

id	category	prompt_text	answer	counterfactual_answer
colors_001	Colors	The color of grass is	green	red
days_001	Days of the week	The day that comes after Sunday is	Monday	Friday
capital_cities_001	Capital cities	The capital of France is	Paris	New Delhi
opposites_001	Opposites	the opposite of hot is	cold	slow
family_relations_001	Family relations	The mother of mother is called	grandmother	grandfather
simple_arithmetic_001	Simple arithmetic	one plus one equals	2	0

---

Citation

@misc{sternberg2026interleavedspeechlanguagemodels,
      title={Interleaved Speech Language Models Latently Work In Text}, 
      author={Talia Sternberg and Gallil Maimon and Yossi Adi},
      year={2026},
      eprint={2606.22473},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.22473}, 
}