Expand description
FastEnhancer-S 48 kHz speech denoiser — a pure-Rust port of the pinned reference.
Reference: aask1357/fastenhancer @ f85223bd546b27f39dc0744e0310dcd246f750a4,
checkpoint release ckpt-v1.0.0-48khz / fastenhancer_s.zip (MIT). The port consumes
the inference-form weights: the reference’s own remove_weight_reparameterizations()
folds every weight-norm and BatchNorm into plain conv/linear weight+bias before export,
so this engine implements only convolutions, GRUs, one tiny frequency attention, and
the compressed-STFT front/back ends.
Geometry (the s config, asserted at load): n_fft 1024, hop 512, 64 encoder channels,
stride-4 frequency downsample (512 -> 128 bins), 3 RNNFormer blocks at 48 channels x 48
frequency slots with 4 attention heads, complex ratio mask output.
Everything is time-causal except the STFT overlap-add; the whole model runs per frame with GRU state carried across frames, so the offline and streaming decompositions are the same arithmetic.
Structs§
- Enhancer
- The complete inference-form parameter set.
Enums§
Constants§
Functions§
- resample_
lanczos6 - Windowed-sinc (Lanczos-6) rate conversion, the same kernel enrollment trusts for its any-rate references: cutoff clamped to the lower Nyquist, taps normalized by their own sum so DC gain stays 1 at the clip edges.
Type Aliases§
- Tensor
Entry - One tensor as handed to
Enhancer::load: shape and row-major data.