CentificAIResearch/audio-loop

Audio Loop: Long-Audio Reasoning Benchmarks

An evaluation set for long-form audio reasoning. It accompanies the paper AudioLoop: Program-Guided Iterative Reasoning Over Long-Form Speech (NeurIPS 2026 Workshop on Long Context Foundation Models), where it is referred to as IQ2-QA.

Config	Task	Audio source	Items
iq2_qa	Multi-hop QA over debates	Recorded live debates	13

Load annotations

from datasets import load_dataset

iq2 = load_dataset("CentificAIResearch/audio-loop", "iq2_qa", split="test")

Each row has a file_name field with the audio path relative to the config root (e.g. "audio/161_Should_the_FTC_Ban_Non-compete_Clauses_.mp3").

Load audio

Audio is not embedded in the parquet stream (files are too large), so file_name is a string. Resolve and decode it explicitly:

from huggingface_hub import hf_hub_download
import soundfile as sf

REPO = "CentificAIResearch/audio-loop"

row = iq2[0]
audio_path = hf_hub_download(
    repo_id=REPO,
    filename=f"iq2_qa/{row['file_name']}",
    repo_type="dataset",
)
audio, sr = sf.read(audio_path)        # numpy array, sample rate
print(audio.shape, sr)

hf_hub_download caches files locally and is a no-op on subsequent calls. To pre-download all audio:

from huggingface_hub import snapshot_download
snapshot_download(REPO, repo_type="dataset", allow_patterns="iq2_qa/audio/*")

---

iq2_qa (IQ2-QA)

Overview

13 complete Oxford-style debate recordings (~50–60 min each) with multi-hop reasoning queries over each debate. In each debate, one or two debaters per side argue FOR and AGAINST a motion, in a live-audience format.

Source

The Open to Debate podcast (formerly Intelligence Squared U.S.): https://podcasts.apple.com/us/podcast/open-to-debate/id216713308. Audio was downloaded from the public RSS feed in April 2026. Questions and reference answers were drafted with Gemini 3 Pro; reference answers are framed in terms of the FOR and AGAINST sides of each motion.

License note

Episodes are publicly distributed for personal listening; included here for non-commercial academic research use only. Annotations are released CC-BY-4.0.

Schema (metadata.jsonl, 1 row per debate)

{
  "file_name": "audio/161_Should_the_FTC_Ban_Non-compete_Clauses_.mp3",
  "debate_id": "iq2_161",
  "file": "161_Should_the_FTC_Ban_Non-compete_Clauses_.mp3",
  "album": "Intelligence Squared U.S. Debates",
  "is_iq2": true,
  "is_debate": true,
  "duration_minutes": 51.3,
  "debate_topic": "Should the FTC Ban Non-compete Clauses?",
  "queries": [
    {
      "query_id": "iq2_161_A",
      "query_type": "claim_rebuttal_counter",
      "query": "What was the main argument made by the FOR side ...",
      "hop_count": 3,
      "requires_acoustic_cues": false,
      "difficulty": "medium",
      "reference_answer": "..."
    }
  ]
}

Query types

Type	Count	Description
claim_rebuttal_counter	13	FOR claim → AGAINST rebuttal → FOR counter (3-hop)
theme_comparison	13	Compare both sides on a shared theme (2-hop)
tone_disagreement	13	Identify the most heated exchange (2-hop, acoustic cues)

39 queries total across 13 debates (3 per debate).

Format

• Audio: MP3 (as published)
• Annotations: JSON Lines, one row per debate

Stats

Stat	Value
Debates	13
Queries	39
Per-debate duration	~50 – 60 min
Total audio runtime	~11 hours

---

Licensing summary

Item	License
iq2_qa/ annotations	CC-BY-4.0
iq2_qa/ audio	Publicly distributed podcast episodes; included for non-commercial academic research

---

Changes

• October 2026: The aae_tts config (AAE-TTS, a two-speaker speech rendering of the Argument Annotated Essays corpus, version 2) was removed. The AAE2 license does not permit redistribution of the corpus, so AAE-TTS is used for evaluation in the paper but is not released.