[ASR] Mask-based dereverb algorithm

[anteju]

What does this PR do ?

This PR adds a multichannel input, multichannel output mask-based dereverberation algorithm based on weighted prediction error (WPE).

Collection: ASR

Changelog

• New: Dereverberation algorithm implemented in MaskBasedDereverbWPE
• Added unit tests for handling of convolution matrices
• Input length is now optional for AudioToSpectrogram
• Updated relevant unit tests

Usage

An example script with a test vector is attached to this PR.

Example script

pr_5693_example_script.zip

Code example

The algorithm can be used as follows:

import soundfile as sf
import torch

from nemo.collections.asr.modules.audio_modules import MaskBasedDereverbWPE
from nemo.collections.asr.modules.audio_preprocessing import AudioToSpectrogram, SpectrogramToAudio
from nemo.collections.asr.parts.preprocessing.segment import AudioSegment

# Parameters
fft_length = 512
hop_length = fft_length // 2
filter_length = 8
delay = 2
num_iterations = 3

# Load audio
sample_rate = 16000
x = AudioSegment.from_file(filename, target_sr=sample_rate).samples.T

# Add batch dimension, shape (B, C, T)
x = x[None, ...]

# Prepare analysis and synthesis transforms
stft = AudioToSpectrogram(fft_length=fft_length, hop_length=hop_length)
istft = SpectrogramToAudio(fft_length=fft_length, hop_length=hop_length)

# Analysis transform
X, _ = stft(input=torch.tensor(x))

# Prepare dereverb instance
dereverb = MaskBasedDereverbWPE(filter_length=filter_length, prediction_delay=delay, num_iterations=num_iterations)
# Processing
Y, _ = dereverb(input=X, mask=None)

# Synthesis transform: shape (B, C, T)
y, _ = istft(input=Y)

# Save audio
sf.write('output/dereverb_output.wav', y.cpu().numpy()[0, ...].T, sample_rate)

Signal example

Below we show spectrograms of the input signal and the processed output signal. Signals are available in the attached example. Only the first channel is shown.

Input signal	Output signal

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

• New Feature
• Bugfix
• Documentation

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

• Related to # (issue)

[jbalam-nv]

LGTM