Sidon v1.0.0

First stable release. Full, tiny and corresponding-source images are published and verified in both official registries.

Sidon restores noisy or degraded speech, separates two speakers, and converts audio through a local browser studio and HTTP APIs. The full Docker image includes pinned model weights and works offline; the tiny image downloads them on first use and retains them in the persistent cache.

• Enhance a single voice at 48 kHz, split a mixture into two 16 kHz tracks, or enhance both separated tracks at 48 kHz. Adjust the original/restored mix.
• Convert without loading a model. Upload common audio formats and export WAV, FLAC, Opus, MP3, AAC/M4A or Ogg Vorbis, individually or in a ZIP with a manifest.
• Use waveform players, retained-job history, persisted defaults, model residency controls and GPU telemetry in the browser studio.
• Integrate native /v1/audio APIs with direct files or asynchronous jobs, structured errors, cancellation and bounded FIFO admission. The API follows familiar OpenAI conventions; enhancement/separation are Sidon extensions.
• Run optimized adaptive F32 restoration on supported NVIDIA GPUs, with folded projections for longer windows and cached fused vocoder kernels.
• Deploy Python 3.13, PyTorch 2.14.1 and CUDA 13.0 with hashed runtime dependencies, checksum-verified weights, component notices and matching source images.

Linux x86-64 GPU deployments require a CUDA 13-compatible host driver and NVIDIA container support. CPU inference is also supported. Persist /app/persistent. For hosted use, set SIDON_API_KEY, use a TLS/access-controlled reverse proxy, and isolate tenants in separate instances. One worker processes jobs with four total admission slots by default; clients retry 429 responses using Retry-After. There is no streaming; clients submit complete audio chunks.

The operator confirmed quality on his tested material after the security upgrade. Shared-host timing observations do not establish a performance regression. Restoration may change timbre or introduce artifacts, and two-speaker labels can swap in ambiguous overlaps. See qualification, security, deployment and licensing.

Project contributions use Apache-2.0 with retained MIT/ISC material. Sidon and MossFormer2 weights retain their publisher grants; container dependencies retain their own terms and corresponding-source requirements.

Deployment references

Full image

docker run -d --name sidon --restart unless-stopped --gpus all -p 127.0.0.1:7868:7868 -v sidon_data:/app/persistent hangrylabs/sidon:v1.0.0@sha256:0125e73d1d09070bbd38323156e89c357a9fc6fa4b7773e85b77651a3f150c8a

Tiny image

docker run -d --name sidon --restart unless-stopped --gpus all -p 127.0.0.1:7868:7868 -v sidon_data:/app/persistent hangrylabs/sidon:v1.0.0_tiny@sha256:ed7ba7f147f92041ebdbd7d5abd1f8264771ffd7df7c0263d14a633f573728f1

Choose one variant and keep the same persistent volume when replacing it. For GHCR, replace hangrylabs/sidon with ghcr.io/hangry-labs/sidon; the verified top-level OCI index digests are identical in both registries.

Image	Top-level OCI index
v1.0.0	sha256:0125e73d1d09070bbd38323156e89c357a9fc6fa4b7773e85b77651a3f150c8a
v1.0.0_tiny	sha256:ed7ba7f147f92041ebdbd7d5abd1f8264771ffd7df7c0263d14a633f573728f1
v1.0.0_sources	sha256:d0f482663af91a55dc24928d6ce977dc087678debb6fc6413256ad473b4ded60

Corresponding sources: hangrylabs/sidon:v1.0.0_sources@sha256:d0f482663af91a55dc24928d6ce977dc087678debb6fc6413256ad473b4ded60. Binary labels also identify their unique matching source build; see source extraction.