| license: apache-2.0 | |
| library_name: coreml | |
| tags: | |
| - coreml | |
| - audio-classification | |
| - audioset | |
| - ced | |
| # CED AudioSet sound-event classifiers — CoreML (coremlit) | |
| CoreML conversions of the four official **CED** (Consistent Ensemble Distillation) AudioSet sound-event taggers, one fp16 `mel → logits` graph per size, for [`coremlit`](https://github.com/findit-studio/coremlit)'s `audio::ced` module. | |
| | size | embed_dim / heads | source | | |
| |---|---|---| | |
| | `ced-tiny` | 192 / 3 | [mispeech/ced-tiny](https://huggingface.co/mispeech/ced-tiny) | | |
| | `ced-mini` | 256 / 4 | [mispeech/ced-mini](https://huggingface.co/mispeech/ced-mini) | | |
| | `ced-small` | 384 / 6 | [mispeech/ced-small](https://huggingface.co/mispeech/ced-small) | | |
| | `ced-base` | 768 / 12 | [mispeech/ced-base](https://huggingface.co/mispeech/ced-base) | | |
| Each `ced-<size>/ced_<size>.mlmodelc` takes a 64-bin log-mel spectrogram and emits **pre-sigmoid** logits over the 527 AudioSet classes (the caller applies the sigmoid). The log-mel front-end runs in Rust (`coremlit`'s `MelExtractor`), so the graph starts at the mel, not the waveform. | |
| ## Provenance | |
| Converted from-official by `coremlit/crates/coremlit/conversion/ced/` — pinned source revisions + per-file SHA-256 in each `ced-<size>/MANIFEST.json` and `coremlit`'s `tests/ced/model_io.rs`. fp32 parity vs the **unmodified PyTorch pre-sigmoid** oracle: cosine `1.0000000`, max|Δlogit| ~1e-5; shipped fp16 ≥ 0.99999. The weight VALUES are unchanged; the stated graph changes are the mel-front-end lift and the dropped final `sigmoid`. | |
| ## License & attribution | |
| Apache-2.0. Upstream: **Xiaomi Corporation — CED** (Dinkel et al., 2023; arXiv 2308.11957; RicherMans/CED). Retain the Apache-2.0 license and copyright notices, and state changes, when redistributing. | |