soundevents-dataset 0.2.0

Typed, zero-allocation Rust access to Google's AudioSet sound-event taxonomy. Two views are available — the full 632-entry ontology and the 527-class rated label set used by released AudioSet models — both baked in at compile time as &'static data, with case-insensitive perfect-hash lookup.

Installation

[dependencies]
soundevents-dataset = "0.2"

By default this pulls in the rated module — the 527-class label set used by released AudioSet/YAMNet/VGGish models. To use the ontology view instead (or in addition), pick the features explicitly:

# Just the full AudioSet ontology, no rated set.
soundevents-dataset = { version = "0.2", default-features = false, features = ["std", "ontology"] }

# Both views.
soundevents-dataset = { version = "0.2", features = ["ontology"] }

Two views, two modules

Module	Source	Entries	Use when…
`rated`	`class_labels_indices.csv`	527	You're working with model outputs / multi-hot label tensors. Each entry carries its `index` so the position in a 527-vector resolves to a name in `O(1)`.
`ontology`	`ontology.json`	632	You need the full taxonomy, including abstract container nodes (`"Human voice"`, `"Music"`, …) and the 105 entries that aren't in the released rated set.

The two are independent: each lives in its own module, has its own &'static consts, its own perfect-hash map, and its own type (SoundEvent vs RatedSoundEvent). Enable only what you need to keep the binary small.

`rated` — AudioSet rated label set (527 entries)

RatedSoundEvent exposes the same metadata accessors as SoundEvent (id, name, description, aliases, citation_uri, children, restrictions) plus a rated-only index() — the integer 0..527 used as the position in released AudioSet models' output vectors. Walking children() stays inside the rated namespace: any ontology child that is not in the rated set is dropped, so the hierarchy remains self-consistent.

Case-insensitive, separator-distinct lookup

from_key is keyed by UncasedStr, so any case form of an alias resolves to the same entry without us having to enumerate every possibility:

Separator styles are still indexed independently ("man speaking" ≠ "man_speaking" ≠ "man-speaking" ≠ "manSpeaking"), so you only pay for the four shapes the codegen actually emits — every case variant of each shape collapses into one phf bucket.

Features

Feature	Default	What you get
`std`	✓	Standard library + `std`-dependent error reporting via `thiserror`. Disable for `no_std`.
`rated`	✓	The `rated` module (527 entries, ~1900 phf keys).
`ontology`		The `ontology` module (632 entries, ~2400 phf keys).
`alloc`		Opt-in `alloc` support for `no_std` targets with an allocator.
`serde`		Derives `Serialize` for `SoundEvent`, `RatedSoundEvent`, and `Restriction`.

The crate is #![no_std]-compatible (default-features = false). The entire dataset lives in &'static memory: no allocations, no startup cost, and the perfect-hash map is generated by phf_codegen at codegen time so the dataset crate's compile graph contains no proc-macros from phf.

Regenerating the dataset

src/ontology/generated.rs and src/rated/generated.rs are checked in and produced from assets/ontology.json and assets/class_labels_indices.csv by an xtask binary. After updating either source file or the codegen logic, regenerate both with one command:

cargo xtask codegen

License

soundevents-dataset is under the terms of both the MIT license and the Apache License (Version 2.0).

See LICENSE-APACHE, LICENSE-MIT for details. Bundled AudioSet metadata attribution and upstream license details are documented in THIRD_PARTY_NOTICES.md.