Noise Reduction using Spectral Gating

Warnings

• breaking The API for `noisereduce` underwent a significant breaking change in version 2.0.0. The primary `reduce_noise` function's signature and behavior changed, and the old API was moved to `noisereduce.noisereducev1.reduce_noise`.
  Fix: Update your code to use the new `noisereduce.reduce_noise` function. If you need to maintain compatibility with older code, import `reduce_noise` from `noisereduce.noisereducev1`.
• breaking Version 3.0.0 introduced a new PyTorch-based implementation for `noisereduce`, offering significant performance improvements and the ability to integrate into neural network architectures. While the original `reduce_noise` function still exists, new PyTorch-specific functionality requires `noisereduce.nn.NoiseReduce`.
  Fix: For performance-critical applications or integration into deep learning pipelines, consider migrating to the `noisereduce.nn.NoiseReduce` module and installing `noisereduce[torch]`.
• gotcha The `stationary` parameter in `nr.reduce_noise` (default: `True`) significantly impacts results. While `True` works well for constant background hums, `False` is crucial for non-stationary noise sources like speech or music, though it can be computationally more intensive.
  Fix: Experiment with `stationary=True` and `stationary=False` based on the characteristics of your noise. For voice or complex environmental sounds, `stationary=False` is generally recommended.
• gotcha As of version 3.0.3, `librosa` is no longer a direct dependency of `noisereduce`. If your existing code relies on `librosa` for audio loading, resampling, or other utilities in conjunction with `noisereduce`, you will need to explicitly install `librosa`.
  Fix: If your project requires `librosa`, ensure it is explicitly listed in your project's dependencies (e.g., `pip install librosa`). For basic audio file I/O, `soundfile` is a common alternative.

Install

• pip install noisereduce Core library
• pip install noisereduce[torch] With PyTorch backend (optional)

Imports

• reduce_noise

  from noisereduce import reduce_noise

• NoiseReduce (PyTorch module)

  from noisereduce.nn import NoiseReduce

• reduce_noise (legacy v1)

  from noisereduce.noisereducev1 import reduce_noise

  Only use this import for the legacy (v1.x) API after upgrading to v2.0.0+ if compatibility is required.

Quickstart

This quickstart demonstrates how to generate a simple noisy audio signal and apply noise reduction using both the default stationary and the more robust non-stationary modes of the `noisereduce.reduce_noise` function. It also includes comments on how to use the optional PyTorch backend for higher performance.

import noisereduce as nr
import numpy as np

# --- 1. Generate dummy noisy audio ---
rate = 44100  # sampling rate
duration = 5  # seconds
t = np.linspace(0, duration, int(rate * duration), endpoint=False)

# Clean signal (e.g., a sine wave)
clean_audio = 0.5 * np.sin(2 * np.pi * 440 * t) # A4 note

# Add some random noise
noise = 0.2 * np.random.randn(len(t))
noisy_audio = clean_audio + noise

# --- 2. Reduce noise ---
# For stationary noise (default and generally faster)
reduced_noise_stationary = nr.reduce_noise(
    y=noisy_audio, 
    sr=rate, 
    stationary=True
)

# For non-stationary noise (e.g., speech with varying background noise)
# This is often more effective but can be slower.
reduced_noise_non_stationary = nr.reduce_noise(
    y=noisy_audio, 
    sr=rate, 
    stationary=False
)

print(f"Original audio shape: {noisy_audio.shape}")
print(f"Reduced audio (stationary) shape: {reduced_noise_stationary.shape}")
print(f"Reduced audio (non-stationary) shape: {reduced_noise_non_stationary.shape}")

# --- Optional: Using the PyTorch backend (requires `pip install noisereduce[torch]`) ---
# try:
#     import torch
#     model = nr.nn.NoiseReduce(sr=rate, nonstationary=False)
#     audio_tensor = torch.from_numpy(noisy_audio).float().unsqueeze(0) # Add batch dim
#     reduced_audio_tensor = model(audio_tensor)
#     reduced_audio_pytorch = reduced_audio_tensor.squeeze(0).numpy()
#     print(f"Reduced audio (PyTorch) shape: {reduced_audio_pytorch.shape}")
# except ImportError:
#     print("PyTorch not installed, skipping PyTorch example.")