pyrnnoise

Python bindings for RNNoise, a recurrent neural network for audio noise reduction.

Features

• Real-time noise suppression for speech audio
• Command-line interface for processing audio files
• Supports mono and stereo audio files
• Visualize voice activity detection probabilities

Installation

pip install pyrnnoise

Usage

Command-line interface

# Basic usage
denoise input.wav output.wav

# With voice activity detection plot
denoise input.wav output.wav --plot

Python API

from pyrnnoise import RNNoise

# Create denoiser instance
denoiser = RNNoise(sample_rate=48000)

# Process audio file
for speech_prob in denoiser.denoise_wav("input.wav", "output.wav"):
    print(f"Processing frame with speech probability: {speech_prob}")

Advanced Usage

The RNNoise class provides several methods for processing audio at different levels:

• denoise_frame(frame, partial=False): Process a single audio frame (480 samples at 48kHz)

  • Returns a tuple of (speech_probabilities, denoised_frame)
  • speech_probabilities: Voice activity detection probabilities for each channel
  • denoised_frame: The denoised audio frame

• denoise_chunk(chunk, partial=False): Process a chunk of audio data

  • Takes a numpy array of audio samples [num_channels, num_samples]
  • Yields tuples of (speech_probabilities, denoised_frame) for each frame
  • Useful for processing audio streams or large audio files in chunks

Example using denoise_chunk:

import numpy as np
from pyrnnoise import RNNoise

# Create denoiser instance
denoiser = RNNoise(sample_rate=48000)

# Generate or load some audio data (stereo in this example)
audio_data = np.random.randint(-32768, 32767, (2, 48000), dtype=np.int16)

# Process audio chunk
for speech_prob, denoised_audio in denoiser.denoise_chunk(audio_data):
    print(f"Speech probability: {speech_prob}")
    # Process denoised_audio as needed

Build from source

# Clone with submodules
git submodule update --init

# Build RNNoise library
cmake -B pyrnnoise/build -DCMAKE_BUILD_TYPE=Release
cmake --build pyrnnoise/build --target install

# Install Python package in development mode
pip install -e .

License

Apache License 2.0