silero_vad

Silero VAD v6 (16 kHz) as a streaming Keras model with explicit LSTM state.

Follows snakers4/silero-vad v6.2.2 (commit 60b7ffa), silero_vad_16k_op15.onnx. One call takes 576 samples, the last 64 samples of the previous call followed by 512 new ones (32 ms at 16 kHz), and returns the speech probability. The LSTM state is carried as state_in_0/state_in_1 (h, c), fed from the previous call’s state_out_0/state_out_1; zeros (and zero context) start an independent stream. Weights come from helia_edge.importers with SILERO_VAD_V6_ONNX (silero_vad_params).

The default options compute the reference model exactly: an STFT magnitude with a square root. The other options (SileroVadParams) compute the same model in other ways: max_projection needs no square root, so the model exports to int16 activations; conv_blocks and live_taps are exact rewrites. They derive what they need (block kernels, the folded padding, tap slices) from the same weights inside the graph, so an exporter folds them into constants.

Machine-readable model

SAMPLESconstantSamples per call: 64 of context and 512 new ones.
UNITSconstant
SileroBlockStftclassSTFT magnitude of one call (576 samples) from convolutions over 64-sample blocks.
SileroFrameConvclassAn encoder Conv1D with ReLU over frames held as rows of a (frames, 1, channels) image.
SileroLiveTapsclassA zero-padded encoder convolution as a dense layer over the kernel taps that see real frames.
buildfunctionBuild the Silero VAD v6 16 kHz streaming model, untrained.

SAMPLES = 576

Samples per call: 64 of context and 512 new ones.

UNITS = 128

SileroBlockStft(magnitude: str = 'sqrt', **kwargs={})

STFT magnitude of one call (576 samples) from convolutions over 64-sample blocks.

The same transform as StftMagnitude(256, 128, 129, padding=(0, 64)) with the same stored basis (256, 1, 258): frames 0 to 2 are a convolution over blocks with stride 2, and the last frame’s right reflect padding is folded into its kernel, so no reflect pad or strided frames over samples remain. The kernels are derived from the basis in the graph and an exporter folds them into constants. The output is (batch, 4, 1, 129), the frames as rows of an image, which the encoder convolves without reshapes.

Parameters

Parameters of SileroBlockStft
Name	Type	Default	Description
`magnitude`	`str`	'sqrt'	``sqrt`` (exact) or ``max_projection`` (no square root, within 0.25%).

magnitude = magnitude

build(input_shape)

call(audio)

compute_output_shape(input_shape)

get_config()

SileroFrameConv(filters: int, strides: int = 1, **kwargs={})

An encoder Conv1D with ReLU over frames held as rows of a (frames, 1, channels) image.

The kernel keeps the Conv1D shape (3, channels, filters); input padding is a separate layer.

filters = filters

strides = strides

build(input_shape)

call(x)

compute_output_shape(input_shape)

get_config()

SileroLiveTaps(filters: int, taps: tuple[int, ...], **kwargs={})

A zero-padded encoder convolution as a dense layer over the kernel taps that see real frames.

The kernel keeps the convolution’s shape (3, channels, filters). taps lists the taps that see the input’s frames, in order; the input is those frames’ channels, flattened.

filters = filters

taps = tuple(taps)

build(input_shape)

call(x)

compute_output_shape(input_shape)

get_config()

build(
    params: SileroVadParams,
    input_shape: tuple[int | None, ...] | None = None,
    *,
    batch_size: int | None = None,
    name: str | None = None,
) -> keras.Model

Build the Silero VAD v6 16 kHz streaming model, untrained.

Inputs are audio (576,) float32 in [-1, 1] and the state state_in_0 and state_in_1 (128,). Outputs are prob (1,) and state_out_0 and state_out_1. Streaming and export use batch_size=1.

Parameters

Parameters of build
Name	Type	Default	Description
`params`	`SileroVadParams`	Required	Model parameters: the geometry fixed by the v6.2.2 weights, and the options.
`input_shape`	`tuple[int \| None, ...] \| None`	None	None, or the audio shape ``(576,)``.
`batch_size`	`int \| None`	None	Static batch size; None for a dynamic batch.
`name`	`str \| None`	None	Model name; the family when None.

Returns

Returns of build
Type	Description
`keras.Model`	keras.Model: The model, named ``silero_vad`` unless ``name`` is given.