# SAID (PaSST) checkpoint card ## Intended use This checkpoint reproduces the SAID (PaSST) model reported in the paper. It predicts separate 180×360 labeled acoustic maps from four Eigenmike capsule signals at 10 fps. ## Identity - Canonical filename: `said_passt.ckpt` - Format: flat exponential-moving-average PyTorch state dictionary - Size: 547,886,573 bytes - SHA256: `85b72bd4a0749dd4ca080ddea30eba68a903af465d916408a03e21912609ca1b` - Runtime tensors: 741 - Submitted tensors covered by the fingerprint: 757 ## Input and output contract The model accepts the **1-based Eigenmike capsule numbers** `[6, 10, 26, 22]` at 48 kHz in two-second windows and predicts 16 candidate slots over the fixed 13-class taxonomy. It produces acoustic maps and class/confidence values; it does not predict source distance. The checkpoint contains Sony and TAU recording-domain embeddings. Sony is the generic inference default and is not automatic domain detection. DCASE evaluation selects the embedding from the dataset subset. ## Reported development-test metrics Here (4.2) means 2026-04-02, commit `84b2cd1`; (6.30) means 2026-06-30, commit `d4df662`. See [Evaluation](../evaluation.md#metric-versions). | Macro mAP (4.2) | Mask AP (4.2) | Macro mAP (6.30) | Class-agnostic AP (6.30) | Macro Pearson r (6.30) | Macro Class-F1 | |---:|---:|---:|---:|---:|---:| | 0.120150 | 0.237972 | 0.124514 | 0.197483 | 0.426790 | 0.388488 | The metrics use all 78 full development-test recordings. Pearson correlation uses the official macro protocol. ## Training provenance ### PaSST initialization The Class Feature Encoder descends from the official AudioSet-pretrained PaSST-S checkpoint: - Upstream repository: - Upstream revision: `2a5c818afcc2a215b2a1aaf1ed8be71f89d43201` - Architecture identifier: `passt_s_swa_p16_128_ap476` - Initial-weight asset: `passt-s-f128-p16-s10-ap.476-swa.pt` - Initial-weight URL: - Initial-weight SHA256: `302903fa8c4aee817b11dc982da0b29aaf8d11a3e722420476d0a12c9db70c2c` - PaSST software license: Apache-2.0; see the [included license](../../LICENSES/PaSST-Apache-2.0.txt) The PaSST encoder was adapted to produce frequency-pooled temporal Class Features, further trained for DCASE sound recognition, and frozen while the complete SAID detector was trained. ### SAID training data - Audio2Sph pretraining used online simulated scenes containing VCTK sources. - Complete SAID training used SourceBank indices over VCTK, MUSDB18-HQ, FSD50K, and verified FSDKaggle2018 clips, rendered into online scenes. - Final fine-tuning used the DCASE2026 Task 3 Track A development-training recordings and labels. The internal SourceBank index contained 122,359 rows and has SHA256 `489f2405025d64f087a76a753cb0fae1051e91df00236ae2285ca2046d664535`. It is retained as provenance and is not distributed. Audio files and dataset labels are not included with the checkpoint. The complete data provenance is summarized in the [training data notice](../../LICENSES/TRAINING_DATA.md). ## Distribution and license The checkpoint is distributed for non-commercial research under the [SAID Model Weights Non-Commercial Research License 1.0](../../LICENSES/SAID-Model-Weights-NonCommercial-1.0.txt), subject to every applicable upstream term. It contains a PaSST-derived backbone and parameters trained from the SourceBank mixture described above. The SAID weight license applies only to rights held by Runbang Wang and does not replace upstream terms. Original training audio and labels are not included. The PaSST project distributes its software under Apache-2.0 and presents the initial-weight asset above as an official pretrained model for inference and fine-tuning. The SourceBank mixture includes non-commercial, restricted, or Sampling+ terms; this release therefore applies the non-commercial research boundary to the complete checkpoint. Exact revisions and attribution are recorded in [Third-party notices](../../THIRD_PARTY_NOTICES.md). Contains information from the CSTR VCTK Corpus, made available under the ODC Attribution License 1.0. The corpus is copyright 2012, Centre for Speech Technology Research, University of Edinburgh, and was constructed by Christophe Veaux, Junichi Yamagishi, and Kirsten MacDonald. The official VCTK 0.80 record is .