AI Samples uses Google Magenta GANSynth acoustic_only, the publisher's official
TensorFlow.js checkpoint. Model origin:
https://storage.googleapis.com/magentadata/js/checkpoints/gansynth/acoustic_only
GANSynth and Magenta.js: Google Inc. and Magenta contributors, Apache-2.0.
See MAGENTA-LICENSE and https://github.com/magenta/magenta-js.
Research: Jesse Engel, Kumar Krishna Agrawal, Shuo Chen, Ishaan Gulrajani,
Chris Donahue and Adam Roberts. GANSynth: Adversarial Neural Audio Synthesis, ICLR 2019.
https://magenta.tensorflow.org/gansynth
Training data: Google NSynth, CC BY 4.0.
https://magenta.tensorflow.org/datasets/nsynth
https://creativecommons.org/licenses/by/4.0/
Exact mirrored model bytes and SHA256 hashes: model/integrity.json.
Runtime: TensorFlow.js 3.21.0, Apache-2.0; fft.js 4.0.4, MIT.
New Tha Spot wrapper: MIT. Dependencies and notices: dependency-manifest.json and licenses/.
Source: /source/creator-studio-addons.tar.gz, vendor/magenta and build-samples.mjs.
Changes to Magenta.js 1.23.1: worker-compatible globals; extraction of four
reconstruction utilities from core/audio_utils.js; default FFT import for bundling.
Audio output is neural generation followed by peak normalization and short fades.
This tool does not replace openDAW's instruments or sampler. Import the WAV into them.

Fixed upstream inverse-STFT crop end index (start + length) to retain the full four-second output.