ltuncay/chime-audioset-strong-balanced

CHIME · AudioSet Strong

This is a source-preserving prepared candidate, not a scientifically admitted CHIME benchmark release. It repackages the selected upstream release into split-specific WebDataset TAR shards with native audio and per-example JSON. A Parquet index permits metadata filtering without downloading audio shards.

Source and selected scope

• Upstream: AudioSet Strong.
• Pinned source release: 0c609e8302cf139307f639c57652032af0a88041.
• Selection: Frozen balanced AudioSet Strong benchmark selection: 19,954 training clips, 1,030 validation clips and all 16,437 usable strong-eval clips. Training and validation share the native train container split; task manifests specify their roles.
• Original license/access conditions: CC BY 4.0 (AudioSet annotations; mirror card). Audio is YouTube-derived and mirror-provided; original uploader rights apply. Public ungated Hugging Face mirror at an immutable commit with per-object LFS SHA-256; Google strong labels from the official AudioSet download page.
• Source notes: Audio from the pinned agkphysics/AudioSet mirror (eval complete; balanced train and 40 evenly spaced unbalanced-train shards for a bounded strong-labelled training pool). Strong labels from Google. Native mirror bytes are decoded to sample-exact WAV at preparation.

Original authors, licensors and recording contributors retain their rights. This preparation was assembled for public noncommercial research access; this repository does not grant broader rights than the original sources. Preserve per-recording attribution and notices. provenance/ contains original annotations/docs/licenses, source provenance and the redistribution audit. FSD50K/Clotho audio has per-recording terms; Clotho captions retain the complete custom Tampere experimental/noncommercial license. The CHIME code license does not apply to these data.

Representation and preservation

There are 37,421 audio assets, 96,227,559,960 stored-audio bytes and 96,225,913,395 decoded-PCM-equivalent bytes. Native sample rate, duration and channel count are retained. No loudness normalization, silence trimming, resampling or downmixing is applied.

Every audio payload uses WAV (including WAVEX/RF64). Existing WAV bytes are retained. Other source codecs are decoded once into sample-exact integer or floating-point WAV; floating-point WAV avoids quantizing MP3/Vorbis decoder output. Every converted frame is read back and its float64 sample digest compared with the original decoder. The source encoding/hash and decoder version remain in each converted record's container_standardization metadata. Sampling rate, frame count, channel count, annotations and sample keys are unchanged. Each payload is accompanied by <key>.json in an ordinary TAR. The JSON contains split, stored audio metadata, provenance, source groups and source targets. manifest.jsonl also records shard/member offsets and checksums. metadata.parquet provides audio and shard columns, common source/group/license columns, and exact arbitrary annotations in metadata_json (decode with json.loads). It does not duplicate audio payloads. Keeping heterogeneous annotations as JSON preserves empty objects and differing target schemas without Arrow type coercion. release.json hashes the audio shards and record manifest, and publication.json inventories all published files including attribution assets.

CHiME-3 background recordings interleave six synchronized native mono PCM16 channels into one six-channel PCM16 WAV per session, with exact decoded-sample checks and an explicit channel map. The excluded CHiME speech corpus is not included. Container standardization preserves decoded samples exactly; source encoded bytes differ for converted assets.

Source splits

• test: 16,437 original audio assets.
• train: 20,984 original audio assets.

These are original source partitions or explicitly unspecified/background partitions. ESC-50 retains five official folds; no fixed benchmark train/test assignment is invented. VocalSet's split is the supplied technique-classification singer protocol, not a universal task protocol. Background noise is not converted into artificial word or silence examples. Linked views/channels share provenance groups.

Scientific limitations

Scientific admission remains pending. Native file integrity and packaging checks do not establish label correctness, perceptual deduplication, contamination freedom or benchmark validity. Existing upstream annotation errata and ambiguity are retained/documented. The source audit found declared Freesound ID overlaps within some official folds/partitions and across ESC-50, FSD50K and Clotho; FSD50K train/validation also share uploaders. Use the supplied source IDs, group metadata and provenance/source-audit.json when defining an evaluation protocol. Do not claim these candidate splits are universally leakage-free. Scores from this repository alone are not official CHIME scores.

Loading

Pin an immutable Hub commit revision in experiments. A generic Hugging Face WebDataset loader can read the shard configurations:

from datasets import load_dataset
samples = load_dataset("ltuncay/chime-audioset-audioset-strong", revision="<commit-sha>", streaming=True)

For local source-native decoding and checksum verification, use CHIME's iter_release(path, split=..., verify_hashes=True). Its audio arrays have shape [frames, channels]; the evaluated model owns its preprocessing. Download only selected shards/index files when appropriate; benchmark execution must not silently fetch data.

Provenance and citation

Cite the upstream dataset and recording authors as specified in the original documentation under provenance/. This preparation does not claim authorship of the source recordings or annotations. Exact upstream objects, checksums, release selection, preparation metadata and source URLs are recorded in release.json. Source Zenodo metadata/citations are retained when available. See the CHIME project implementation brief and scientific paper for the evolving evaluation protocol.

CHIME v1 benchmark usage

Do not train on CHIME evaluation data: 16437 records in the release split(s) test, listed one key per line in provenance/benchmark-v1/eval-keys.txt. Models submitted to CHIME must not see these recordings (or their source recordings) during pretraining or fine-tuning.

Task	Family	Metric	Folds	Train	Validation	Test (eval)
audioset-strong-coarse-sed	frame_multilabel	segment_f1	1	train	validation	test
audioset-strong-sed	frame_multilabel	segment_f1	1	train	validation	test
audioset-strong-speech-music	frame_multilabel	segment_f1	1	train	validation	test
audioset-strong-tagging	multilabel	macro_ap	1	train	validation	test

Exact per-record partition assignments (including records excluded from training because they share a source with evaluation data, and digitally silent assets) are in provenance/benchmark-v1/<task>.jsonl; the frozen task definitions are in provenance/benchmark-v1/usage.json. Splits never depend on the evaluation seed.

Benchmark subset release

This repository contains only the frozen CHIME benchmark selection. The larger prepared source release remains available at https://huggingface.co/datasets/ltuncay/chime-audioset-audioset-strong/tree/42c10b1efd38d66bb5035b715eb98a296a0e2992. The internal dataset family ID is retained for task compatibility. AudioSet tagging protocol version 2 shares the balanced SED training and validation selection. Prior full-training tagging scores are not comparable. The complete usable strong-eval test selection is preserved.