This Python library reduces substantial background noise in audio files containing speech. It uses a machine learning model (38m parameters) trained to handle different types of ambient noise.
pip install audio-denoiser
In Windows, you need the soundfile audio backend:
pip install soundfile
In Linux, both the soundfile and sox audio backends should be supported. Note that the library is trained with the soundfile backend.
Basic:
from audio_denoiser.AudioDenoiser import AudioDenoiser
denoiser = AudioDenoiser()
in_audio_file = '/content/input-audio-with-noise.wav'
out_audio_file = '/content/output-denoised-audio.wav'
denoiser.process_audio_file(in_audio_file, out_audio_file)
With additional options:
from audio_denoiser.AudioDenoiser import AudioDenoiser
import torch
import torchaudio
# Use a CUDA device for inference if available
device = torch.device('cuda:0') if torch.cuda.is_available() else torch.device('cpu')
denoiser = AudioDenoiser(device=device)
in_audio_file = '/content/input-audio-with-noise.wav'
out_audio_file = '/content/output-denoised-audio.wav'
auto_scale = True # Recommended for low-volume input audio
denoiser.process_audio_file(in_audio_file, out_audio_file, auto_scale=auto_scale)
You can also provide your own waveform tensor:
from audio_denoiser.AudioDenoiser import AudioDenoiser
noisy_waveform, sample_rate = torchaudio.load('/content/input-audio-with-noise.wav')
denoiser = AudioDenoiser()
denoised_waveform = denoiser.process_waveform(noisy_waveform, sample_rate, auto_scale=False)
print('Tensor shape: ', denoised_waveform.shape)