Inference runtime notices

The default Sidon wheel/source distribution includes the studio/API, restoration wrapper and selected MossFormer2 SS architecture. Its source license expression is Apache-2.0 AND MIT AND ISC. This fork’s own contributions use Apache-2.0, as set out in the root LICENSE. Installed dependencies, CUDA libraries and separately downloaded model weights retain their own terms.

Sidon

Copyright (c) 2026 Hangry Labs contributors. New integration code and this fork’s own contributions are offered under Apache-2.0, in the root LICENSE. This project-level choice does not relicense third-party material or replace its retained notices.

ClearerVoice-Studio / MossFormer2 SS

The four model files under sidon/inference/_mossformer2 are adapted from ClearerVoice-Studio revision 6b3774dc79c46ae8bed2a4fa5f706f0ac8c75c61: https://github.com/modelscope/ClearerVoice-Studio

Authors credited upstream include Shengkui Zhao, Zexu Pan and Trung Hieu Nguyen. The architecture is offered under Apache-2.0. The original text is included as licenses/ClearerVoice-Apache-2.0.txt in the source distribution and under the wheel’s .dist-info/licenses directory.

Changes made on 2026-10-29: separation-only packaging; removal of unused imports; replacement of the channel-layer-normalization wrapper with PyTorch LayerNorm and explicit axis transposes. The original third-party layer_norm.py, training, metrics, video and other model families are not included. Source hashes and URLs for each imported file are recorded in _mossformer2/provenance.json.

The model implementation credits SpeechBrain’s dual-path implementation (Apache-2.0) and FLASH-pytorch (MIT). Their license texts are preserved in licenses/SpeechBrain-Apache-2.0.txt and licenses/FLASH-MIT.txt, respectively, and in the wheel’s license directory. FLASH’s notice is:

Separately downloaded weights

Model: alibabasglab/MossFormer2_SS_16K

Publisher model card, declaring Apache-2.0: https://huggingface.co/alibabasglab/MossFormer2_SS_16K/blob/407cb030cd66340918ebb6c8cc63b18f8592cdbe/README.md

Revision: 407cb030cd66340918ebb6c8cc63b18f8592cdbe

File: last_best_checkpoint.pt

SHA-256: 00a3a48bda492db1e829b85dd443f8f43a43039a3e90f1a24962ea9caf14a11a

Checkpoint bytes are verified before loading; loading uses PyTorch’s weights_only=True with strict parameter matching. Model weights are downloaded separately into the Hugging Face cache and are not embedded in the wheel/sdist.

The full Docker image bakes in this pinned asset for offline inference. Its distribution must include the model license and attribution alongside the application and dependency notices; see docs/docker-packaging.md.

Binary dependencies

Python package licenses do not replace the terms on bundled native libraries. In particular, SoundFile wheels can bundle LGPL-licensed libsndfile and related codec libraries, while PyTorch GPU builds include NVIDIA components. Preserve their installed notices and meet the applicable source/relinking and NVIDIA redistribution requirements before distributing a complete binary environment. These components allow commercial use subject to their terms; they are not all licensed under the source package’s MIT/Apache expression.

The repository contains only the selected inference implementation. The upstream research code, old demo, training/data pipelines and experimental dependency group have been removed. See docs/licensing.md for the current source review and docs/mossformer-runtime.md for runtime validation.

Sidon restoration weights and feature preparation

Publisher: sarulab-speech. Model repository sarulab-speech/sidon-v0.1, revision b3b02d8bbd55fdbc410e6e46e76ef95ace4fbf52, publisher-declared MIT. The pinned publisher card is preserved as licenses/Sidon-MODEL-CARD.md. The original Sidon MIT grant and attribution are in licenses/Sidon-upstream-MIT.txt.

CPU and CUDA feature_extractor_*.pt / decoder_*.pt TorchScript exports are loaded directly, without training/model Python code or flow-matching. Exact file hashes are recorded in sidon/inference/assets.py and the baked image’s model manifest; every asset is verified before loading. TorchScript is executable model content, so only the pinned, reviewed publisher exports are accepted.

Feature preparation follows upstream Sidon’s Kaldi-filterbank normalization and frame stacking under its MIT grant. This fork adds bounded overlapping inference, right context, silence handling, original-gain restoration and an optional dry/wet mix. These modifications are not claimed to reproduce the upstream demo exactly. Publisher model licenses are not independent guarantees about training provenance.

Studio and build helpers

The shared shell/audio styles, GPU monitor, source/notice collectors and publication planner are adapted from faster-qwen3-tts-studio, MIT. Its complete original notice is retained in licenses/faster-qwen-upstream-MIT.txt. New Sidon integration uses Apache-2.0. The packaged Lucide subset carries its own complete license at sidon/web/static/vendor/lucide/LICENSE. See THIRD_PARTY_NOTICES.md for the supplemental Intel oneMKL notice and docs/docker-packaging.md for corresponding sources.

The separation wrapper now follows upstream active-region input normalization and whole-track output RMS calibration. Raw model gain is not a usable playback level. Each separated track is matched to mixture RMS, not an estimate of the original speaker’s loudness; no peak limiter is applied.

Optimized F32 restoration

On supported Linux CUDA systems, restoration can bind the pinned exports into an inference-only PyTorch implementation. The automatic path preserves the reference encoder/arithmetic for shorter windows and uses a folded encoder with strict F32 math for windows of at least 512 latent frames. All windows use a fused F32 Snake activation in the vocoder. The folded encoder merges 16 LoRA branches once at model load and projects relative-position embeddings before gathering attention bias. Weights, normalization, chunking and output rates are preserved.

Tensor bindings adapt audio.cpp’s Sidon converter at revision 1f5dc02257379b790fae9f3a0e0d28065d9baf2f, copyright 2026 ShugoAI LLC, Apache-2.0; its complete notice is in licenses/audio.cpp-Apache-2.0.txt. The original TorchScript path remains available, including CPU and unsupported platform fallback. Forced native mode uses strict IEEE F32 for every window. Precision flags are restored afterwards; service inference is serialized.

Triton uses the installed C compiler for a local launcher. Compiler/header notices and corresponding source are included with the image’s dependency records. Cached kernels contain compiled program code, not audio.

Audio conversion

FFmpeg and ffprobe are installed as Debian runtime packages for additional upload containers and Opus, MP3, AAC/M4A and Vorbis exports. The application invokes them as separate processes; their GPL/LGPL component licenses remain in force. The runtime inventory retains their Debian copyright notices and exact source identities, and the matching source image includes their corresponding sources. Native Python installs use an operator-installed FFmpeg rather than a bundled executable. No conversion codec or model weights are added to the Python wheel.