[SOLVED] Python: How to separate out noise from human speech in audio file?

Python: How to separate out noise from human speech in audio file?

I have an audio file in which I attempted to filter out the noise:

frequency, array = read('sample/OSR_us_000_0014_8k.wav')
b, a = signal.butter(5, 1000/(frequency/2), btype='highpass')
filteredSignal = signal.lfilter(b, a, newSound)

This highpass filter doesn't appear to be that effective though. Are there better ways to accomplish what I am attempting to do? Additionally, I would prefer to have the extracted background noise. Most algorithms available filter out the noise, but I'd like to extract the noise out as a numpy array.

Solution

There is a deep learning-based neural network pretrained model available in Python for noise removal from audio files. I've used it, and it provides very high accuracy.

However, it only supports WAV audio files with a sampling rate of 48k. If your WAV file has a different sampling rate, you can convert it to 48k using the librosa library..

Steps to Use DeepFilterNet for Enhancing Noisy Audio Files

If your audio file has a different sampling rate, follow these steps:

Install librosa & soundfile library:

pip install librosa soundfile

Use the following function to resample the audio:

import librosa
import soundfile as sf

def re_sample_audio(audio_path, sr):
    print("Resampling audio")
    audio, _ = librosa.load(audio_path, sr=16000)
    audio_resampled = librosa.resample(audio, target_sr=sr, orig_sr=16000)
    sf.write(audio_path, audio_resampled, 48000)
    print("Resampling done")

re_sample_audio(audio_path, 16000)

Installing Necessary Libraries for DeepFilterNet

Install the required PyTorch version (>=1.9) with CPU or CUDA support from pytorch.org, e.g.:

pip install torch torchaudio -f https://download.pytorch.org/whl/cpu/torch_stable.html

Install DeepFilterNet:

pip install deepfilternet

Enhancing the Noisy Audio File

Use the following code to enhance a noisy audio file:

from df.enhance import enhance, init_df, load_audio, save_audio
from df.utils import download_file

if __name__ == "__main__":
    # Load the default model
    model, df_state, _ = init_df()

    # Path to the noisy audio
    audio_path = "noisy_audio_path.wav"

    audio, _ = load_audio(audio_path, sr=df_state.sr())

    # Denoise the audio
    enhanced = enhance(model, df_state, audio)

    # Save the enhanced audio
    save_audio("enhanced.wav", enhanced, df_state.sr())

Alternatively, You Can Use the Command Line to Enhance Noisy Audio Files

# Specify an output directory with --output-dir [OUTPUT_DIR]
deepFilter path/to/noisy_audio.wav