Silero server VAD for the engine-resident duplex session.
The detector keeps the split upstream's serving-side server_vad module used, because the split is what makes one model serve many sessions: a backend scores one 512-sample frame and is otherwise stateless, and the per-stream state -- the partial frame, the model state, the endpoint counters -- belongs to :class:SileroStreamingVAD, one per session.
Two backends. :class:SileroVADBackend runs the pinned Silero v6.2 ONNX graph on CPU and is shared process-wide, which is why its state is passed in and out rather than held on the session object. :class:TorchSileroBackend loads the same model through the silero-vad package for environments that have no local ONNX artifact; torch keeps its state inside the module, so that one is per session.
Endpointing follows Silero v6.2's streaming hysteresis: activation uses the configured threshold, while a turn can only end on frames below max(threshold - 0.15, 0.01). Once a silence candidate exists, louder frames keep the elapsed-silence clock running but cannot themselves close the turn.
module-attribute ¶
module-attribute ¶
Bases: RuntimeError
property ¶
scratch_bytes: int
Bytes this detector holds between chunks (the partial frame only).
process_base64(
audio: object, *, fmt: object, sample_rate_hz: object
) -> StreamingVADResult
Drop stream state while preserving the session-relative audio clock.
The clock survives so speech_start_ms after a barge-in still refers to the session's timeline; _stream_start_ms then stops the prefix padding reaching back into audio that was discarded.
Shared ONNX Runtime Silero v6.2 detector running on CPU.
Resolve, verify and load one detector backend per engine process.
ONNX is preferred and shared; the torch package is the fallback and is built per call because its state is internal.
get() -> SpeechDetectorBackend
The shared ONNX backend, or a per-session torch one when it cannot be built.
An explicitly configured server_vad_model_path never falls back: the operator named that artifact, so a missing file or a missing ONNX Runtime is their error, not something to paper over with a different model.
dataclass ¶
dataclass ¶
The same model through the silero-vad package, for hosts with no ONNX artifact.
Torch keeps the recurrent state inside the module, so unlike the ONNX backend this one cannot be shared: each session gets its own instance and new_state only resets it.