Other
SAID / docs /checkpoints /audio2sph_l.md
IN03X's picture
Release Audio2Sph-L 543k weights and document four-second SourceBank generation
177e832 verified
|
Raw History Blame Contribute Delete
2.26 kB
# Audio2Sph-L checkpoint
Released on 2026-11-11 as an extension to the paper models.
- File: [`audio2sph_l.ckpt`](https://huggingface.co/IN03X/SAID/blob/main/audio2sph_l.ckpt).
- Training step: 543,000; original multi-loss supervision, not the BCE-only variant.
- Input: four channels, 48 kHz, four seconds.
- Output: Any and 13 class-union maps at 45 × 90, plus 128-channel spherical features.
- Format: tensor-only state dictionary, 550 tensors, 147,306,278 bytes.
- SHA256: `d4992ee47c971526bae5499e6272370526565a30ecee50ececdc589cf432df99`.
This initializes Audio2Sph-L or the encoder of SAID-L, not the paper SAID
architecture. No pretrained complete SAID-L checkpoint is supplied.
See [Audio2Sph-L](../audio2sph_l.md) for inference and training.
## Data and rights
The historical training source pool includes VCTK, MUSDB18-HQ, FSD50K,
FSDKaggle2018, ASA2 stems and AudioSet-indexed recordings, with measured
TAU room responses/backgrounds and additional FSD50K non-target interference.
The public data recipe uses the four directly downloadable SourceBank corpora
and the same five scene families; it does not reconstruct the full historical
source pool. No source recordings are distributed with these weights.
License: [SAID Model Weights Non-Commercial Research License 1.0](../../LICENSES/SAID-Model-Weights-NonCommercial-1.0.txt).
Copyright (c) 2026 Runbang Wang. The license grants only rights held by the
licensor and preserves applicable third-party terms.
Source attributions are supplied in
[`provenance/audio2sph_l_source_attributions.tsv`](https://huggingface.co/IN03X/SAID/blob/main/provenance/audio2sph_l_source_attributions.tsv).
ASA2's dataset card declares MIT, but does not provide a complete constituent
recording rights chain. AudioSet annotation licensing does not establish
rights to the indexed YouTube recordings. These provenance limitations are
disclosed; a non-commercial release is not a certification of all third-party rights.
We acknowledge the Audio Research Group at Tampere University for the
[TAU Spatial Room Impulse Response Database](https://zenodo.org/records/6408611).
Its [full notice](../../LICENSES/TAU-SRIR.txt) permits experimental,
non-commercial use and prohibits commercial distribution of results.