AuraBench Dataset Card

Summary

AuraBench contains 48 canonical audio questions rendered under 16 deterministic signal-processing image conditions. Every condition uses the same source waveform, question, options, and deterministic answer key. Three additional controls remove the image, replace it with a uniform blank image, or permute mel images within each task family.

The shuffled-image control uses a deterministic semantic derangement within each compatible choice set: no image remains on its source question, no image crosses to a different option vocabulary, and every source image's ground-truth label differs from the target question's answer. Its taker-visible exam identity matches the ordinary Mel condition, so the model is not told that image labels have been permuted.

Evaluated models receive image and text inputs only. The Any-to-Bench bundles do not contain audio files, transcripts, source filenames, corpus labels, or descriptive alt text.

Tasks

Family	Questions	Labels
Speaker category	8	corpus-provided female / male label
Language identification	8	English, German, Spanish, French, Mandarin Chinese, Japanese, Korean, Arabic
Sound-event classification	8	dog, rain, footsteps, keyboard, clock, helicopter, siren, glass
Phonetic content	8	four English vowels and two complete Mandarin aspiration pairs
Lexical recognition	8	isolated English digits
Emotional expression	8	happy, sad, angry, fearful, crossed with two actors

All questions are single-choice and worth one point. Grading is deterministic.

Public question IDs are opaque random identifiers. They carry no family, source, or ordinal meaning, and display numbers are omitted. Within each family, question order is shuffled; every question has its own pinned option permutation; and correct-option positions are counterbalanced for comparable choice sets. The same audited public layout is used across every image condition to preserve paired evaluation.

Evaluation integrity

The published Any-to-Bench bundles include grading.json so results can be reproduced. That file must never be staged into the taker's workspace. For agentic CLI takers, a private HOME, disabled web search, blocked shell network, and Codex workspace-write mode are still insufficient by themselves: workspace-write permits reads elsewhere on the host, including the source bundle and previous run outputs.

The reference AuraBench runner therefore applies an outer OS-level filesystem policy. The taker can read and write its current ephemeral a2b run root and can read the pinned Codex runtime, but it cannot read the project directory, other a2b workspaces, mounted volumes, or network mounts, and cannot write outside its current run root. OS canaries verify these boundaries before benchmark execution. See scripts/hermetic-bin/codex and scripts/run_matrix.

The corrected reference matrix contains 38 condition-effort pairs and 76 model runs, all with 48/48 coverage and zero schema errors. The hidden shuffled-Mel control scores 7/48 and 10/48 for sol low/medium, compared with the former impossible 48/48 result obtained when host answer keys were readable. Detailed results and SHA-256 digests are in results/summary.json.

Image conditions

Atomic conditions are waveform, linear STFT, mel spectrogram, CQT, multi-resolution STFT, reassigned spectrogram, wavelet scalogram, ERB cochleagram, periodicity maps, modulation energy, source/filter tracks, and a prosody/voice-quality dashboard. Combination conditions are waveform+mel, mel+periodicity, a speech atlas, and a general acoustic atlas.

All transformations are classical DSP. AuraBench does not use ASR, neural audio encoders, neural classifiers, learned embeddings, neural denoisers, or neural source separation.

Source corpora and licenses

Corpus	Used for	Upstream license
RAVDESS speech	speaker category, emotion	CC BY-NC-SA 4.0
FLEURS	language identification	CC BY 4.0
ESC-50	sound events	CC BY-NC 3.0; per-clip attribution remains applicable
Hillenbrand et al. vowels	vowel classification	hosted with permission in an MIT-licensed repository
Yali-Lower	Mandarin consonants	GPL 3.0 or later
Free Spoken Digit Dataset	spoken digits	CC BY-SA 4.0

The corpus licenses remain applicable to canonical audio and derived images. This repository's MIT license covers AuraBench code and original metadata only; it does not replace upstream data licenses. See data/provenance/audio_manifest.json for the source ID, citation, license string, source hash, and canonical hash of every item.

Preprocessing

Each source is decoded to mono, DC-centered, resampled to 16 kHz with a polyphase filter, peak-normalized to 0.98, center-cropped only when longer than the family target, and right-padded with zeros when shorter. Canonical files are PCM16 WAV. Target lengths are 1 second for phonetic and digit items, 4 seconds for speaker and emotion items, 5 seconds for sound events, and 10 seconds for language identification.

No sample was selected or removed based on a model's performance. FLEURS transcript fields are deliberately discarded during ingestion.

Known source properties

• RAVDESS emotional expressions are acted speech rather than spontaneous emotion.
• Speaker-category ground truth is the source corpus's binary metadata label; it is not a claim about gender identity inferred from voice.
• ESC-50 documents that some upstream clips may contain class-dependent preprocessing.
• Hillenbrand items are /hVd/ syllables rather than isolated steady-state vowels.
• A representation may be undefined for a signal. Unreliable F0 and formant estimates remain blank rather than being replaced by a fallback estimate.

Bundle layout

Each directory under benchmark/bundles/ is a complete Any-to-Bench bundle containing exam.json, answer_schema.json, grading.json, manifest.json, aurabench.json, and any required image assets. aurabench.json records the image-to-question mapping, including the deliberate permutation used by the shuffled-image control, using only opaque public IDs. config/public_layout.yaml pins the reproducible question and option permutations and records the answer-shortcut audit score.