# DeepFilterNet 3 for johnmcm.com

The WebAssembly speech-enhancement engine behind johnmcm.com's Remove
Background Noise page (/remove-background-noise).

It is [DeepFilterNet](https://github.com/Rikorose/DeepFilterNet) by Hendrik
Schröter, whose code is dual-licensed MIT OR Apache-2.0 (both licence files
are served beside the build). The pretrained DeepFilterNet3 model ships in
the same repository and is compiled into the author's own MIT/Apache-labelled
library by default; no separate licence for the weights is stated. It was
trained on the DNS Challenge 4 data (CC BY 4.0 and other open licences).

What is built, from commit `d375b2d8309e0935d165700c91da9de862a99c31`:

- `libDF`, the Rust runtime (STFT, ERB features, the tract ONNX engine and
  deep filtering), with wasm-pack for ES-module workers, WebAssembly SIMD on,
  and binaryen's `wasm-opt -O3`.
- `src/wasm.rs` replaces `libDF/src/wasm.rs`: the model is passed in rather
  than compiled in; the thresholds are those of the upstream `deep-filter`
  command line tool (-15/35/35 dB), which keep speech at its level (the
  runtime defaults leave it about 5 dB quieter); a whole chunk is processed
  per call; and the delay (STFT plus look-ahead) is reported so the page can
  line the output up with the input.
- `src/worker.mjs`: the worker the page talks to. It loads the wasm and the
  model once, cleans each piece it is sent, and trims the delay.

`build.ps1` runs the Docker build (Rust image pinned by digest, crates by the
repository's Cargo.lock) and copies the result to
`frontend/vendor/local/deepfilter@1/`, with the model renamed
`deepfilternet3.bin`. A shipped `@N` folder must not change: bump it, and the
path in `src/app/remove-background-noise/noise.ts`, before rebuilding.