fastenhancer

FastEnhancer folded-inference model for streaming spectral enhancement.

Adapted from aask1357/fastenhancer (revision e74cab1, models/fastenhancer/ default/model.py), the revision of the onnx-vd-v1.0.0 release. One call processes one STFT frame: spec_in is (n_fft // 2 + 1, 1, 2) real/imag, and each RNNFormer block carries a GRU state (freq, channels). Callers zero the states only at independent sequence starts and feed each state_out_k back as state_in_k. The graph matches the released ONNX inference form, with BatchNorm and weight normalization folded. STFT/iSTFT framing, weight files and weight terms are the caller’s concern.

Machine-readable model

linear_filterbanksfunctionReturn fixed triangular (pre, post) frequency projections as [in, out] kernels.
FastEnhancerCompressionclassDrop the Nyquist bin and compress magnitude: x max(|x|, 1e-5)^(c - 1).
FastEnhancerMaskOutputclassApply a complex mask, undo compression and restore a zero Nyquist bin.
FastEnhancerFrequencyProjectionclassProject the frequency axis of (batch, freq, channels) with a [in, out] kernel.
FastEnhancerGRUStepclassOne GRU step per frequency band, gates z, r, h and reset after matmul.
FastEnhancerPositionalEmbeddingclassAdd a learned (freq, channels) embedding.
FastEnhancerFrequencyAttentionclassMulti-head self-attention across frequency bands, no output projection.
buildfunctionConstruct an untrained one-frame FastEnhancer with named streaming inputs and outputs.

linear_filterbanks(n_freq: int, n_filter: int) -> tuple[np.ndarray, np.ndarray]

Return fixed triangular (pre, post) frequency projections as [in, out] kernels.

Follows upstream revision e74cab1, which produced the onnx-vd-v1.0.0 weights; upstream changed this formula in 8d2d419 (2026-02-25).

FastEnhancerCompression(compression: float, **kwargs={})

Drop the Nyquist bin and compress magnitude: x * max(|x|, 1e-5)^(c - 1).

compression = compression

call(spec)

get_config()

FastEnhancerMaskOutput(compression: float, **kwargs={})

Apply a complex mask, undo compression and restore a zero Nyquist bin.

compression = compression

call(inputs)

get_config()

FastEnhancerFrequencyProjection(units: int, **kwargs={})

Project the frequency axis of (batch, freq, channels) with a [in, out] kernel.

units = units

build(input_shape)

call(x)

get_config()

FastEnhancerGRUStep(units: int, **kwargs={})

One GRU step per frequency band, gates z, r, h and reset after matmul.

Matches ONNX GRU with linear_before_reset=1 and PyTorch nn.GRU: h’ = z * h + (1 - z) * tanh(Wx + bw + r * (Rh + br)). Returns h’.

units = units

build(input_shape)

call(inputs)

get_config()

FastEnhancerPositionalEmbedding()

Add a learned (freq, channels) embedding.

build(input_shape)

call(x)

FastEnhancerFrequencyAttention(num_heads: int, use_bias: bool = False, **kwargs={})

Multi-head self-attention across frequency bands, no output projection.

The qkv kernel is [channels, 3 * channels] with per-head interleaved columns: head h uses q = 3dh*h + [0, dh), k = + [dh, 2dh), v = + [2dh, 3dh).

num_heads = num_heads

use_bias = use_bias

build(input_shape)

call(x)

get_config()

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

Construct an untrained one-frame FastEnhancer with named streaming inputs and outputs.

Inputs: spec_in (bins, 1, 2) and state_in_k (freq, channels) per block. Outputs: spec_out and state_out_k. Fixed linear filterbanks are initialized and frozen; other weights are untrained until imported with a mapping from fastenhancer_params (FASTENHANCER_T_ONNX or fastenhancer_mapping).

Parameters

Parameters of build
Name	Type	Default	Description
`params`	`FastEnhancerParams`	Required	Model parameters.
`input_shape`	`tuple[int \| None, ...] \| None`	None	None, or ``spec_in``'s fixed shape ``(bins, 1, 2)``.
`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 ``fastenhancer`` unless ``name`` is given.