aTENNuate is a network that can be configured for real-time speech enhancement on raw audio waveforms. It can perform tasks such as audio denoising, super-resolution, and de-quantization. This repo contains the network definition and a set of pre-trained weights for the aTENNuate model.

Note that the repo is meant for denoising performance evaluation on custom audio samples, and is not optimized for inference. It also does not contain the recurrent configuration of the network, so it cannot be directly used for real-time inference by itself. Evaluation should ideally be done on a batch of .wav files at once as expected by the denoise.py script.

Please contact Brainchip Inc. to learn more on the full real-time audio denoising solution. And please consider citation our work if you find this repo useful.

One simply needs a working Python environment, and run the following

pip install attenuate

To run the pre-trained network on custom audio samples, simply put the .wav files (or other format supported by librosa) into the noisy_samples directory (or any directory of your choice), and run the following

from attenuate import aTENNuate

model = aTENNuate()
model.from_pretrained("PeaBrane/aTENNuate")
model.denoise('noisy_samples', denoised_dir='denoised_samples')

# denoised_samples = model.denoise('noisy_samples')  # return torch tensors instead

The denoised samples will then be saved as .wav files in the denoised_samples directory.