Privacy & licenses

Updated 1 November 2026. Nocito Music is free. It has no accounts, advertising trackers, paid inference or hidden cloud generation.

Your audio

Browser arrangement decodes and processes the selected file on your device. Audio is not uploaded to Nocito hosting. The original file is not overwritten. Results remain temporary until you download them or explicitly save them in this browser's local library (IndexedDB, up to 500 MiB). Clearing site storage removes saved copies. Language preference is stored locally.

Local AI is optional. When you click Connect, a separate localhost window asks for permission. Only after you approve can this page send your chosen audio and settings to the ACE-Step engine on your computer. The page can retrieve and cancel only jobs from that connection. It cannot browse an existing private studio library or obtain its token. Close or disconnect the local window to revoke access and cancel unfinished jobs from that connection.

The companion retains original uploads, generated WAV files, source permission declarations, prompts, seeds and reports in its own data folder. Disconnecting does not delete them. You control this local folder. The source-only installation kit contains no personal music, model weights, Python runtime or binary libraries.

Network requests

Hosting providers process normal web requests (including IP address, requested URL and browser information) to deliver the site. No audio is included in those requests. The optional installer downloads Python from python.org, runtime packages from PyPI, AMD or PyTorch, and pinned model files from Hugging Face. These providers receive normal download requests. Generation runs offline after setup. We do not promise that a browser, OS or a separately installed program makes no network requests.

Optional support and recommendations

Buy Me a Coffee is an external, voluntary support link. All Music results and downloads remain free. No payment scripts are loaded; visiting the support page is your choice. Links to Nocito Expense Tracker and Reminder include campaign parameters identifying Nocito Music as the referring tool. They contain no audio, prompt, file name or user identifier.

Licenses and provenance

The Nocito website and companion source use the MIT license. The kit includes a manifest of its source files and exact upstream revision. ACE-Step source is MIT at commit ca1e85fe9430179831e6bc6be790c332190a3866. The official model card declares MIT at model revision 19671f406d603126926c1b7e2adc169acbcade22; that repository has no separate weights license file. The Qwen3-Embedding-0.6B text encoder declares Apache-2.0. The installer retains downloaded model cards, checksums and dependency license files locally.

Runtime dependencies have individual licenses. These include LGPL-3.0 frozendict, MPL-2.0 certifi and LGPL-2.1 libsndfile. Their installed notices and provenance are retained by the installer. Binary environments and model weights are not redistributed in this kit.

Only use recordings you own or are permitted to adapt. A source declaration records your permission; it does not independently verify a third-party license. No guarantee is made about copyright protection, motif preservation, similarity, Content ID claims or suitability of each generated passage for commercial use.

Practical limits

Browser arrangement can make an output of up to ten minutes by reusing parts of the source, with crossfades. It does not compose new music or promise seamless joins. Export is 48 kHz float32 WAV; sample-rate conversion cannot improve a compressed source. The level meter reports sample peak and RMS, not integrated loudness (LUFS).

Local AI is a Windows x64 beta. The measured machine is an RX 6600 8 GiB / Ryzen 5600G / 48 GiB RAM using an experimental AMD runtime. Native ten-minute generation took 742.8 seconds with peak framework-reserved VRAM 6.58 GiB and process RAM 15.98 GiB. This is a hardware-specific benchmark, not a speed guarantee. Other NVIDIA devices are an unmeasured installation path following upstream CUDA pins; setup must pass its own real generation test. CPU-only generation, other AMD GPUs, macOS and Linux are not validated by this kit.

Source variations use five settings of the same Turbo model and produce thirty-second excerpts. Extend a chosen result using arrangement. AI may change melody or add digital artifacts. No new model variants, stem separation or LoRA training are included.

References: official ACE-Step project, official model repository, Qwen text encoder.