Native Silero VAD model on Lstm with weights from the upstream ONNX file #1278

Context

Silero VAD (v6.x, MIT) is the de-facto on-device voice activity detector. A SKaiNET consumer currently has to ship ONNX Runtime only for it, next to whatever runs the ASR model, which costs a second native runtime, an extra ~2.3 MB model on a separate loader, and a minSdk floor set by ORT rather than by SKaiNET.

The model is small: an STFT front end baked into the graph as conv weights, a conv encoder, one LSTM cell with a carried [2,1,128] state, and a dense sigmoid head. Input per call is a 64-sample context prepended to a 512-sample frame at 16 kHz. Lstm with LstmState(h, c) and an explicit step already exists in skainet-lang-core (#823), which was added with exactly this kind of streaming use in mind.

Scope

• SileroVad module (skainet-models or a new skainet-model-vad): the graph expressed with ctx.ops.* only, so it runs eagerly and lowers through the tape path like every other SKaiNET model. Reuse Lstm.
• Weights loaded from the initializers of the pinned upstream silero_vad.onnx through skainet-io-onnx's weight reader (no new file format; upstream stays the source of truth). Load the STFT basis as weights rather than re-deriving it so the model matches bit for bit.
• Streaming API: process(frame: FloatArray): Float returning speech probability, reset(), and the state carried across calls; a VoiceActivityDetector-shaped adapter for the audio-side libraries.
• Parity test: probabilities within 1e-4 of ONNX Runtime over ≥ 1000 consecutive frames from three clips including state carry-over, checked against a committed golden (ORT is a test-time dependency only).
• Android measurement: < 1 ms per frame on one thread with the JNI tier.
• Ground-truth entry so the Python-side suite (Ground truth: expand Python-side op coverage to match OperationExecutor #985) covers the op set.

Non-goals

• Importing the Silero graph through the ONNX graph importer. That needs LSTM/Gather/Unsqueeze/Where/If converters and is tracked separately as the verification path; the hand-written module is the shipping path.

Related

• [Feature]: LSTM layer (Lstm + explicit-state step API) in skainet-lang-core #823 — Lstm layer with explicit state step API
• ONNX I/O module MVP: #230 — ONNX I/O module MVP