Voxa.Audio.Onnx 0.7.2-alpha

Voxa.Audio.Onnx

The shared ONNX Runtime session host for Voxa's local-speech tier (VLS-006). Every ONNX model in Voxa — VAD, TTS, ASR, enhancement — loads its weights through this host instead of constructing an InferenceSession itself, so a model loads once per (path, device) and is shared across every connection on the process.

This package ships the host + the device seam + the model-descriptor shape — not a model. Consuming engines bring their own pinned catalogs (a ParakeetCatalog, a SidonCatalog, …) the same way Kokoro and whisper.cpp do.

What's in the box

• OnnxModelHost — a process-wide (path, device)-keyed session cache. Load(path, device, hook) returns a shared IOnnxSession; concurrent loads of the same key resolve to one instance. Sessions are process-lifetime (they hold shared weights); OnnxModelHost.EvictAll() disposes and clears them for tests and Studio's "unload models."
• IOnnxSession — a thin, test-fakeable handle exposing the InferenceSession (for engines that bind OrtValues directly — the zero-allocation steady-state pattern), the input/output names, and the EP that actually loaded.
• OnnxDevice + OnnxDeviceParser — the shared Device config convention (cpu / auto / cuda / directml / coreml), parsed identically across every ONNX engine.
• OnnxModelDescriptor + ResolveAsync — an ONNX model self-describes as a graph artifact + pinned sidecars (tokenizer / vocab / config / extra graphs) and resolves through the unchanged VoxaModelCache.

CPU by default; GPU is opt-in and never bundled

The base package references only the CPU Microsoft.ML.OnnxRuntime (pinned to the same 1.26.0 as SileroVad / Kokoro, with PrivateAssets so it isn't re-flowed onto consumers). Device=cpu is the default and the only target the bundled runtime supports.

To use a GPU, the consuming app adds the matching ORT package itself — Voxa never ships GPU natives, and the ORT CPU/GPU packages conflict if both land in one process:

<PackageReference Include="Microsoft.ML.OnnxRuntime.Gpu" Version="1.26.0" /> 

With that present, Device=cuda (or directml / coreml) loads the matching EP. Without it, an explicit GPU device fails at session creation with a copy-pasteable remediation, and Device=auto falls back to CPU with a warning — it never hard-fails.

Not yet here

• The OnnxTensors convenience run-helper (the KokoroTtsEngine.RunInference shape, lifted) lands with the first consumer that needs it, once its dtype matrix is known. Today engines use the direct Session tier.
• VLS-006 ships no model catalog; pinned models (and their SHA-256s + cleared licences) live in the consuming engine packages (VLS-004 / 005 / 007 / 008, VRT-005).