Introduction

StaleyNET just hit a same-day open music drop: GitHub v0.1.2 published 2026-11-04T08:14:16Z, staleynet-audio 0.1.3 on crates.io the same morning, and weights at arpanpathak/staleynet-checkpoints (card last touched through the afternoon). It turns a short caption — optionally with lyrics — into a waveform. Every stage is first-party Rust: spectrogram, RVQ tokenizer, Transformer with hand-derived gradients, decoder, Griffin-Lim vocoder, and GPU kernels via NVIDIA cuda-oxide. No pretrained backbone.

Trained on an NVIDIA Jetson Orin NX (16 GB) and Jetson AGX Thor (128 GB) — edge boxes, not a cloud fleet. The pitch is measurement: because every stage is owned, each bottleneck can be named.

What Shipped

  • Vendor / product: StaleyNET — Arpan Pathak
  • Modality: generative text-to-music (caption ± lyrics → waveform)
  • Open vs closed: code Apache-2.0; checkpoints CC BY-SA 4.0 (subject to Song Describer / JamendoLyrics upstream terms)
  • Runnable today: cargo add staleynet-audio for the audio stack; clone the GH repo for training / CUDA path; download HF run folders for weights
  • Concrete fact: text-only runs are 7,099,904 params; lyrics runs 7,165,440; pq-large is 42,858,496 params on a band-group tokenizer (4 band groups × 4 levels = 16 codebooks of 1024); audio features are 128 mel bands @ 24 kHz (FFT 1024, hop 256)
  • Companion artifacts: preprint PDF + mdBook, dataset arpanpathak/staleynet-data (~3.1 GB)

Run Surfaces

cargo add staleynet-audio

The crate page walks decoding, resampling, log-mel, and Griffin-Lim with five examples that run as documentation tests. Full generation against a checkpoint still lives in the GitHub repo (cuda-oxide Git deps keep the CUDA crates off crates.io for now). Pull a run such as thor-base, lyrics-thor, lyrics-thor-long, or pq-large from the HF checkpoints card, then follow the repo’s generate path.

Caption style is short and literal — e.g. "grunge rock, raw male voice" — with optional lyrics conditioning on the lyrics-trained runs.

Design Notes

Training: audio → 24 kHz log-mel → residual k-means / band-group tokens → delayed staircase for the Transformer conditioned on caption (+ lyrics) bytes. Generation: sample tokens → reconstruct mel → Griffin-Lim for phase. v0.1.2 organises GPU kernels by topic (matmul, attention, loss, optimizer, DSP…), revises the paper with hardware / tokenizer appendices, and publishes the newer Thor / pq-large artifacts.

Limits Worth Knowing

  • This is a research / edge-training stack, not a Suno competitor on day one — Griffin-Lim and ~7–43M-param transformers set a clear ceiling the authors document.
  • CUDA path depends on cuda-oxide from Git; Orin NX and AGX Thor are the stated build targets.
  • Dataset licenses ride along — respect Song Describer and JamendoLyrics terms when redistributing outputs or further training.
  • Zero prior artrealmai.com coverage; distinct from Sonilo / Audio8 / Suno MCP already on the site.

Original Source

Conclusion

If you want a same-day open text-to-music line you can audit stage-by-stage — Rust end-to-end, Jetson-trained, CC BY-SA checkpoints — StaleyNET v0.1.2 is the November 4 music drop to bookmark, with eyes open on Griffin-Lim limits.

—Anabel ♡