b4ph/audio-sampling-budget-results

Audio sampling-budget study: per-query results, condition metadata, and timing

This repository contains the results of a disclosed custom crop experiment (track VERIFIED_CUSTOM): per-query retrieval outcomes, condition metadata, and matched GPU timing measurements for four audio-text models under center-crop audio budgets of 2/5/10 seconds versus native loading.

This is NOT an official MTEB benchmark result. It is a custom sampling experiment that supplies cropped audio to otherwise pinned model/task pipelines; it makes no claim of official-protocol comparability and must not be listed as ordinary leaderboard scores.

No source audio, no audio crops, no embeddings, no caption text, and no model weights are distributed here (rights review; see Licensing). Download instructions for every source artifact are provided instead.

Reference disclosures

• MSCLAP: native reference is corrected_native_reference_msclap_44100_v1 (protocol amendment v1.2). The unchanged upstream MTEB wrapper writes 48 kHz WAVs and calls msclap with resample=False, while msclap 1.3.4 config expects 44,100 Hz — a rate/timeline mismatch. The correction resamples the decoded waveform to the package-configured rate before cropping; all else unchanged. Results are relative to this disclosed correction, not to the unchanged upstream wrapper.
• MuQ: native reference is compatibility_reference_muq_numpy_bridge_v1 (amendment v1.4). The unchanged wrapper fails before inference (NumPy assigned into a Torch buffer); the adapter changes only that assignment.
• Clotho replaced SoundDescs (amendment v1.3): SoundDescs is missing required corpus item d-nhu05046028 and carries an unresolved BBC permission question. These results are Clotho results, not SoundDescs results.

Data fields

Aggregates and uncertainty (in evidence/analysis/): bootstrap_primary.json (24 primary comparisons: model, task, budget, point diff, 95% CI, Bonferroni one-sided lower bound, quality-pass flag, cluster counts), summary.json (per-condition aggregates with official five-decimal equal-query means and unrounded means as separate fields; Spearman ordering), timing_summary.json (per model/task/condition: five block log ratios, speedup point estimate, descriptive and Bonferroni-corrected Student-t bounds, pass flags, GPU telemetry).

Timing block records (in evidence/timings/): per model — batch size, frozen score-blind subset hash, counterbalanced block schedule, warmup count, per-condition block timings, and full GPU environment records.

Licensing (in licensing/): machine-readable license-count histograms for Clotho audio, SongDescriber audio, and model/task weights. Aggregate counts only; no track-identifying records.

Per-query retrieval outputs (query ID, per-query metric value, top-k IDs and scores, hit vector, relevant IDs) are retained in the private study repository and are available on request; they are not distributed here pending a caption and query-text rights review.

Metric definitions

• SongDescriberT2ARetrieval: recall_at_5 — pytrec_eval recall_5, official equal-query mean, scale 0–1.
• ClothoT2ARetrieval.v2: hit_rate_at_5 — pytrec_eval success_5, official equal-query mean, scale 0–1. Every Clotho qrel has relevance 1; success@5 equals hit rate at cutoff 5.

Uncertainty method

Paired cluster bootstrap over saved per-query outputs: 100,000 resamples, seed 20260905, CPU-only. Resampling unit = source recording (706 SongDescriber track clusters via qrels; 1,045 Clotho recording clusters via repairs/clotho_resolution/verified_query_provenance.json); equal-query weighting preserved within each resample; corpus fixed (corpus-fixed inference). One-sided lower bounds use the Bonferroni quantile 0.05/24. Quality pass rule: corrected lower bound strictly greater than −0.005. Timing: 5 matched counterbalanced blocks per model/task; point = exp(mean log-ratio); one-sided Student-t bound (descriptive α=0.05, 4 df; corrected α=0.05/24); speed pass requires point AND corrected bound > 1.10×.

Results summary

All 24 primary quality comparisons FAIL the −0.005 margin (bootstrap_primary.json). Spearman ordering vs native is 1.0 at all center budgets on both tasks (summary.json). Timing: MSCLAP passes 1.10× everywhere (5.6–8.2× SongDescriber, 1.6–2.3× Clotho); MuQ passes everywhere (2.9–4.8×); larger CLAP passes only SongDescriber 5s/10s (~1.3×); clap-htsat-unfused fails everywhere (Clotho ~1.0×: fixed 10 s native window). Full tables: evidence/ in the study repository's publication package.

Known limitations

• Corpus-fixed, seed-conditional bootstrap: no generalization claim beyond the evaluated corpora.
• Four-model panel: ordering evidence is coarse (Spearman granularity 0.8).
• Center crops change both duration and selection policy vs native random truncation (crop-location confounding); start/end rows are exploratory.
• MSCLAP and MuQ baselines are disclosed corrected/compatibility references.
• The LAION wrappers process a fixed 10-second window regardless of supplied duration, so "seconds supplied" ≠ "seconds processed" for those models.
• SongDescriber timing stratification is degenerate (661/706 tracks ≈ 120 s).
• Training-data overlap between models and corpora is unknown and untested. No contamination or decontamination claim is made.
• No end-to-end wall-clock speed claim; the measured boundary is crop + native preprocessing + transfer + audio encoder forward.

Licensing and redistribution

• Clotho corpus audio is a mix: CC0 415, CC BY 3.0 431, CC BY-NC 3.0 154, Sampling+ 1.0 45 (verified item-level inventory). Clotho captions carry Tampere University's experimental/noncommercial grant.
• SongDescriber corpus audio: per-track CC mix (BY-NC-SA 287, BY-SA 150, BY-NC-ND 195, BY 37, BY-NC 15, BY-ND 14, Art Libre 8 of 706 tracks).
• Model weights: LAION Apache-2.0; msclap code MIT, weights MS-PL/MS research terms; MuQ-MuLan-large weights CC-BY-NC-4.0 (noncommercial).

Accordingly, no source audio, crops, embeddings, captions, or weights are redistributed. To reproduce, obtain the artifacts from their sources:

Artifact	Source
SongDescriber corpus/queries/qrels	dukesun99/SongDescriber-T2A rev c58ad9b08343e56ce412d4da41accd15bffdbd6d (upstream: Zenodo 10072001)
Clotho corpus/queries/qrels	lxercode/clotho_t2a_v2 rev 13cf5105f96e1844ee3b0b84cfba9103a7ebe6a2 (upstream: Zenodo 4783391, validation split)
clap-htsat-unfused, larger_clap_music_and_speech	Hugging Face laion/... (Apache-2.0)
msclap 2023 weights	GitHub microsoft/CLAP rev c47d441165daa21986ead0850660917636a81775 (MS-PL terms)
MuQ-MuLan-large	Hugging Face OpenMuQ/MuQ-MuLan-large (CC-BY-NC-4.0)

Citation and attribution obligations of the original sources (Clotho notice, Tampere caption grant, SDD citation) remain in force for anyone using this package's instructions to obtain and use those artifacts.

Reproduction

python study_runner/run_bootstrap_primary.py   # bootstrap_primary.json
python study_runner/run_analysis.py            # summary.json
python study_runner/summarize_timings.py       # timing_summary.json
python study_runner/make_figures.py            # figures

Dataset acquisition + offline verification gate:

uv venv .venv-data --python 3.11
uv pip sync --python .venv-data/Scripts/python.exe --require-hashes requirements-data.lock
.venv-data/Scripts/python.exe -u data_stage.py --data-root F:/AI/datasets/audio-sampling-budget \
  --report-dir F:/AI/projects/audio-sampling-budget/repairs/clotho_data_stage

Citation

If you use these results, cite the upstream research question (MTEB issue #5362), the SongDescriber dataset (Manco et al., arXiv:2311.10057), and Clotho (Zenodo 4783391), and state clearly that this is a custom crop experiment, not an official benchmark score.