MusGU+ Evaluation: PLAUD

Training runs on a single consumer-grade GPU; in the tested configuration the synth stage fits within a T4-class 16GB card. CPU-only training is not supported, as the training code assumes CUDA.

The system is explicitly designed for small, personal datasets on the order of minutes to a few hours of audio, intended to model a single artist or sound world rather than a large corpus.

Complete pathways are provided to train from scratch on the user's own audio, covering the synth, the grouped-VQ codec, the transformer prior, and export, all driven by a shared configuration file. No pretrained base checkpoints are provided, so every model is trained from the ground up.

A beginner-oriented Colab notebook covers the full training-to-export pipeline alongside an annotated configuration template, but there is no graphical training interface, and setup still requires basic familiarity with Colab, file paths, and running command-line stages. The tooling is research-grade and under active change.

Released under CC BY-NC 4.0, which permits redistribution of trained models and checkpoints with attribution for non-commercial use. Sharing is explicitly allowed but commercially restricted.

A Max4Live device is provided, making the instrument playable inside Ableton Live through a graphical interface. It runs via the nn~ external, which must be available in Max. There is no standalone application or web interface.

The source code is openly available with no paywall, login, or usage limit, and a pretrained example model (drum loops trained on CC0 Freesound material) is provided, so inference is possible out of the box without training first.

Exported models are designed for real-time use in nn~ via a streaming, KV-cached prior, and run faster than real time on consumer hardware in tested configurations. Heavier synthesis options increase CPU cost, and the streaming prior and optional post-net add some latency.

A Max4Live device integrates the model directly into Ableton Live, and through nn~ it can also be used in the Max/MSP and Pure Data visual programming environments alongside other audio and control objects.

Under CC BY-NC 4.0 the material may be used, shared, and adapted for non-commercial purposes with attribution. Commercial use of generated audio is not permitted, and attribution is required.

The public repository has GitHub Issues available (enabled by default) as a developer-oriented contact channel. There is no dedicated musician-facing community space such as a forum or chat server.

The system is primarily audio-domain. An encoder maps input audio to a control trajectory for resynthesis, and the generative prior is steered by a learned style code, an optional territory selection, and slow control envelopes. There is no MIDI or symbolic conditioning.

The synth consumes per-frame control trajectories (loudness, spectral centroid, and latent timbre), and the generative prior is guided over time by control-rate envelopes, a style position, sampling temperature, and guidance strength. Fine detail from the prior is produced autoregressively rather than dictated frame by frame.

Loudness and spectral centroid are exposed as explicit, interpretable per-frame control channels. Timbre is carried by VAE latents and a learned style code that are not individually interpretable, and the style and timbre pathways can interact.

A broad set of runtime parameters is exposed, including a style position, sampling temperature, classifier-free guidance strength, control envelopes, and synth-level attributes such as waveshaping and spectral shaping, together with direct access to the per-frame control and latent representation.