Convolution Reverb

A Python package for applying convolution reverb to audio files using PyTorch. This package provides an efficient implementation of convolution reverb using FFT-based convolution.

Installation

Install the package using pip:

pip install convolution-reverb

Requirements

• Python >= 3.8
• PyTorch >= 2.0.0
• torchaudio >= 2.0.0

See requirements.txt.

Usage

Basic Usage

from convolution_reverb import apply_reverb
import torchaudio

# Load audio files
audio_path = "path/to/your/audio.wav"
ir_path = "path/to/your/impulse_response.wav"
output_path = "output.wav"

# Apply reverb
original, convolved, sample_rate = apply_reverb(
    audio_path=audio_path,
    ir_path=ir_path,
    output_path=output_path,
)

Working with Tensors Directly

If you're already working with audio tensors in PyTorch:

import torch
import torchaudio
from convolution_reverb import apply_reverb

# Load a sample audio, impulse response
audio_wav, audio_wav_sr = torchaudio.load("path/to/your/audio.wav")
ir_tensor, ir_sr = torchaudio.load("path/to/your/impulse_response.wav")

# Apply reverb
original, convolved, sr = apply_reverb(
    audio_wav=audio_tensor,
    audio_wav_sr=audio_wav_sr,
    ir_wav=ir_tensor,
    ir_wav_sr=ir_sr,
    output_path=output_path,
)

API Reference

apply_reverb

apply_reverb(
    audio_path: Union[str, None] = None,
    audio_wav: Union[torch.Tensor, None] = None,
    audio_wav_sr: Union[int, None] = None,
    ir_path: Union[str, None] = None,
    ir_wav: Union[torch.Tensor, None] = None,
    ir_wav_sr: Union[int, None] = None,
    output_path: Union[str, None] = None,
    use_partitioned: bool = True,
    block_size: Optional[int] = None,
    normalize: bool = False
) -> Tuple[torch.Tensor, torch.Tensor, int]

Parameters:

• audio_path: Path to the input audio file
• audio_wav: Input audio as a torch.Tensor (n_channels, n_samples)
• audio_wav_sr: Sampling rate of the input audio tensor
• ir_path: Path to the impulse response file
• ir_wav: Impulse response as a torch.Tensor (n_channels, n_samples)
• ir_wav_sr: Sampling rate of the impulse response tensor
• output_path: Path where the output audio will be saved
• use_partitioned: If True, uses partitioned convolution (overlap‑add) to avoid potential numerical precision issues with large FFT blocks. Recommended for long audio signals.
• block_size: Block size (in samples) for partitioned convolution. If not provided and use_partitioned is True, defaults to 10 seconds of audio.
• normalize: Whether to normalize the output audio

Returns:

• Tuple containing:
  • Original audio waveform (torch.Tensor)
  • Convolved audio waveform (torch.Tensor)
  • Sample rate (int)