noisereduce Documentation

The noisereduce package provides two main approaches to noise reduction via spectral gating:

1. Stationary Noise Reduction: Typically requires an explicit noise signal to calculate statistics. It performs noise reduction over the entire signal uniformly. Use this when you have a clear sample of the noise or when the noise characteristics do not change significantly over time.

2. Non-stationary Noise Reduction: Dynamically estimates and reduces noise concurrently with the signal. Use this when the noise profile changes over time (e.g., varying background sounds).

Noisereduce uses spectral gating to reduce noise. It computes a spectrogram and estimates a noise threshold (gate) for each frequency band to create a mask that gates noise.

There are two primary algorithms:

1. Stationary Noise Reduction: Keeps the estimated noise threshold constant across the entire signal. This is useful when the noise profile does not change over time. It can optionally take a specific y_noise clip to calculate statistics.
2. Non-stationary Noise Reduction: Continuously updates the noise threshold over time. This is useful when noise characteristics change (e.g., background environmental noise). It is motivated by Per-Channel Energy Normalization (PCEN) and is effective when signal events occur on specific timescales.

Install the noisereduce package using pip to perform noise reduction on time-domain signals like speech, bioacoustics, and physiological signals.

pip install noisereduce

The project provides Makefile commands to synchronize the local data/ directory with an Amazon S3 bucket using the aws s3 sync command. This is useful for managing datasets used by the library.

Sync local data to S3

Use make sync_data_to_s3 to recursively upload files from your local data/ directory to an S3 bucket. The destination path will be s3://[your-bucket]/data/.

Sync S3 data to local

Use make sync_data_from_s3 to recursively download files from s3://[your-bucket]/data/ to your local data/ directory.

To use the TorchGate functionality, ensure you have the following Python packages installed:

• librosa
• noisereduce
• soundfile
• torch (required for the TorchGate module)

The reduce_noise function is the primary API for performing noise reduction on NumPy arrays. By default, it performs non-stationary noise reduction. To perform stationary noise reduction, set stationary=True.

from scipy.io import wavfile
import noisereduce as nr

# load data
rate, data = wavfile.read("mywav.wav")

# perform noise reduction
reduced_noise = nr.reduce_noise(y=data, sr=rate)

wavfile.write("mywav_reduced_noise.wav", rate, reduced_noise)

For PyTorch-based workflows, you can use TorchGate (imported as TG). This allows the noise reduction algorithm to be used as a standalone module or integrated into a larger neural network architecture. It supports GPU acceleration via the device parameter.

import torch
from noisereduce.torchgate import TorchGate as TG
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")

# Create TorchGating instance
tg = TG(sr=8000, nonstationary=True).to(device)

# Apply Spectral Gate to noisy speech signal
noisy_speech = torch.randn(3, 32000, device=device)
enhanced_speech = tg(noisy_speech)