audio-clean-booster

The recommended pipeline is no-reference:

MossFormer2 speech enhancement + residual denoise + built-in post-processing

You only need the noisy WAV. A clean reference is optional for advanced experiments.

Install

pip install "audio-clean-booster[clearvoice]"

Use

acb clean noisy_16k.wav noisy_16k_final.wav

To keep the intermediate MossFormer2 output:

acb clean noisy_16k.wav noisy_16k_final.wav --keep-mossformer noisy_16k_mossformer2.wav

Python

from audio_clean_booster import clean_with_best

clean_with_best(
    "noisy_16k.wav",
    "noisy_16k_final.wav",
)

Compare Chunks

acb compare \
  --source noisy:noisy_16k.wav \
  --source final:noisy_16k_final.wav

Open compare_chunks.html.

Advanced Commands

The lower-level commands remain available for experiments:

acb mossformer noisy.wav mossformer.wav
acb deepfilter noisy.wav deepfilter.wav
acb reference-match mossformer.wav reference.wav final.wav --mode strict

For all optional backends:

pip install "audio-clean-booster[all]"