PaperlayerView exact revision on arXiv

Source-linked AI summary

BeepBank-500: A Synthetic Earcon Mini-Corpus for UI Sound Research and Psychoacoustics Research

Mandip Goswami

arXiv:2509.17277v1eess.AScs.SD

TL;DR

BeepBank-500 addresses the need for compact, controllable earcon corpora with a reproducible synthetic dataset and accompanying baselines. The baselines report waveform-family classification accuracy and f0-regression error and tolerance results; the authors also state scope and realism limitations.

  • Problem

    Researchers need compact, controllable earcon datasets for prototyping and psychoacoustic work, while existing large datasets emphasize environmental audio and speech.

    Source: §1 · p. 1

  • Method

    BeepBank-500 provides a compact, scripted synthetic earcon corpus with parameterized tone generation, metadata, and reproducible baselines.

    Source: §1 · p. 1, §3.1 · p. 2

  • Results

    Waveform-family classification test accuracy is 81.1%; on single-tone f0 regression, MAE is 63.66 Hz, MedAE is 0.22 Hz, and 80.2% fall within ±1 semitone.

    Source: §5 · p. 3

  • Takeaways & Limitations

    The corpus is intended for research and education on earcon design, timbre features, and simple robustness testing, including to small reverbs.

    Source: §6 · p. 4

    The dataset prioritizes compactness and controllability over ecological breadth, and synthetic signals may not capture perceptual subtleties of human-designed earcons.

    Source: §7 · p. 4, §6 · p. 4

Abstract

from arXiv · show

We introduce BeepBank-500, a compact, fully synthetic earcon/alert dataset (300-500 clips) designed for rapid, rights-clean experimentation in human-computer interaction and audio machine learning. Each clip is generated from a parametric recipe controlling waveform family (sine, square, triangle, FM), fundamental frequency, duration, amplitude envelope, amplitude modulation (AM), and lightweight Schroeder-style reverberation. We use three reverberation settings: dry, and two synthetic rooms denoted 'rir small' ('small') and 'rir medium' ('medium') throughout the paper and in the metadata. We release mono 48 kHz WAV audio (16-bit), a rich metadata table (signal/spectral features), and tiny reproducible baselines for (i) waveform-family classification and (ii) f0 regression on single tones. The corpus targets tasks such as earcon classification, timbre analyses, and onset detection, with clearly stated licensing and limitations. Audio is dedicated to the public domain via CC0-1.0; code is under MIT. Data DOI: https://doi.org/10.5281/zenodo.17172015. Code: https://github.com/mandip42/earcons-mini-500.

1 Motivation and Scope

BeepBank-500 fills a need for compact, controllable earcon datasets with a small, reproducible, rights-clean synthetic corpus. It is intended for rapid research and teaching, not clinical or safety-critical use.

1 Motivation and Scope

  • BeepBank-500 targets the gap in compact, controllable earcon corpora by spanning key parameters with a few hundred items, scripted generation, full metadata, and frictionless reuse.The audio is CC0-1.0 and the code is MIT-licensed.Source: §1 · p. 1
  • The corpus is intended for rapid UI earcon classification, timbre similarity or embedding analysis, robustness studies, onset detection, and teaching or benchmarking.Source: §1 · p. 1
  • Speech, music, affective labeling, clinical applications, and complex room acoustics beyond lightweight Schroeder reverberation are out of scope.Source: §1 · p. 2

2 Related Resources (Context Only)

BeepBank-500 focuses on compact, fully scripted parametric UI tones, complementing broader environmental audio corpora. The context given is that existing synthetic tone and earcon resources can be scattered, unclear in licensing, or larger than needed for didactic work.

2 Related Resources (Context Only)

  • Synthetic tone banks and earcon papers are scattered across HCI and audio ML venues.Source: §2 · p. 2
  • Some existing assets are not centralized with a DOI, lack a clear license, or are larger than necessary for didactic tasks.Source: §2 · p. 2
  • BeepBank-500 complements broader environmental corpora with a compact, fully scripted recipe focused on parametric UI tones.Source: §2 · p. 2

3 Generation Protocol

The generation protocol combines configurable tone synthesis and lightweight room effects, then samples a deterministic parameter grid. Files use a fixed audio format and stable filename-hash splits.

3.1 Signal Chain

  • The signal chain applies an oscillator, optional amplitude modulation, an ADSR envelope, RMS normalization, and optional Schroeder-style reverberation.Available oscillator families include sine, square, triangle, and two FM variants; the design also varies f0, duration, envelope, and AM settings.Source: §3.1 · p. 2
  • RMS normalization targets nominally −20 dBFS with a hard cap at −1 dBFS to avoid clipping.LUFS may be computed for analysis but is not used for normalization.Source: §3.1 · p. 2

3.2 Parameter Grid

  • The parameter grid crosses waveform family, f0, duration, envelope, AM rate and depth, chord type, and reverb kind before deterministic sampling of 300–500 items.Table 1 summarizes the parameter grid.Source: §3.2 · p. 3, Table 1 · p. 3

3.3 File Format and Splits

  • Audio is mono, 48 kHz, 16-bit PCM WAV, with train, validation, and test splits assigned by hashing filenames for stable partitions.Source: §3.3 · p. 3

4 Metadata Schema and Measures

The metadata table describes individual clips with signal and spectral features, including explicitly flagged proxies. Version 1.0.0 contains 400 clips with deterministic 80/10/10 train, validation, and test splits.

4 Metadata Schema and Measures

  • Each metadata.csv row describes one clip and includes simple spectral statistics and proxies for roughness or inharmonicity.The proxy status is explicitly flagged, and LUFS is optional and blank when not computed.Source: §4 · p. 3
  • Re-computation scripts are included for the metadata features.Source: §4 · p. 3
  • Version 1.0.0 contains 400 clips split train/validation/test at 80/10/10 by deterministic filename hash.Source: §4 · p. 3

5 Baselines and Example Analyses

The paper provides minimal baselines for quick comparisons and signal-diversity checks, alongside a metadata schema and reproducibility materials.

Waveform-family classification

  • 81.1% test accuracy is reported for waveform-family classification using pooled 64-bin log-mel features and logistic regression.The baselines are intended to verify signal diversity and facilitate quick comparisons.Source: §5 · p. 3

f0 regression

  • MAE = 63.66 Hz, MedAE = 0.22 Hz, and 80.2% within ±1 semitone summarize YIN’s f0 regression on 111 single-tone items.The error distribution is heavy-tailed, consistent with occasional octave or subharmonic errors under FM and reverberant or AM conditions.Source: §5 · p. 3, Table 3 · p. 4

Reproducibility

  • Scripts, seeds, and exact dependencies are provided, and baseline output JSON files capture metrics for papers.Source: §5 · p. 3

Metadata

  • Table 2 summarizes the metadata.csv columns, with units embedded in names for clarity.Source: Table 2 · p. 4

6 Ethics, Licensing, and Intended Use

The paper specifies reuse terms and intended uses, while bounding the dataset’s scope and claims.

Licensing

  • Generated audio is dedicated to the public domain under CC0-1.0, and the code is MIT-licensed.Source: §6 · p. 4

Intended use

  • The dataset is intended for research and education on earcon design, timbre features, and simple robustness testing, not safety-critical alerts or clinical purposes.Source: §6 · p. 4

Known limitations and risks

  • Synthetic signals may miss perceptual subtleties of human-designed earcons; reverb is schematic, and psychoacoustic measures are proxies unless explicitly computed.Source: §6 · p. 4

7 Limitations and Future Work

The dataset prioritizes compactness and controllability over ecological breadth; proposed future additions broaden its spatial, acoustic, and evaluative coverage.

Limitations

  • Compactness and controllability are prioritized over ecological breadth.Source: §7 · p. 4

Future work

  • Possible additions include HRTF-based spatialization, more envelopes and FM indices, measured room impulse responses, preference data, and expanded f0 sets with micro-variations.Source: §7 · pp. 4–5

Example analyses

  • Figure 1 presents example log-mel spectrograms across waveform families and AM settings.Source: Figure 1 · p. 5

Availability and Citation

Users of BeepBank-500 are asked to cite the dataset DOI and data note; a plain-text citation is available in the repository’s CITATION.cff.

Availability and citation

Reproducibility Checklist

The checklist specifies deterministic data generation, documented metadata and fixed-preprocessing baselines, licenses, and versioning; the CLI commands cover setup, generation, and running both baselines.

Reproducibility Checklist

  • Deterministic seeds are provided for data generation, and the metadata schema is documented in Section 4 and shipped as CSV.Minimal baseline scripts use fixed preprocessing and emit JSON metrics.Source: Reproducibility Checklist · p. 5
  • Audio uses CC0-1.0 and code uses MIT, with licenses indicated in files and records.The project uses semantic version tags, such as v1.0.0, with CHANGELOG entries.Source: Reproducibility Checklist · p. 5

A Quick Start (CLI)

  • The CLI setup creates and activates a virtual environment, then installs dependencies from requirements.txt.The documented commands use python -m venv and pip install.Source: A Quick Start (CLI) · p. 5
  • The CLI generates audio and metadata with a specified output directory, metadata path, seed, and target count, then runs waveform classification and f0 regression.The generation command specifies seed 13; both baseline commands use the generated audio directory and metadata CSV.Source: A Quick Start (CLI) · pp. 5–6

B Minimal BibTeX for the Dataset

The minimal BibTeX entry identifies the dataset by author, title, year, version, and DOI.

B Minimal BibTeX for the Dataset

Loading 2509.17277v1…