Listening to Sounds of Silence for Speech Denoising

Introduction

This is the repository of the "Listening to Sounds of Silence for Speech Denoising" project. (Project URL: here) Our approach is based on a key observation about human speech: there is often a short pause between each sentence or word. In a recorded speech signal, those pauses introduce a series of time periods during which only noise is present. We leverage these incidental silent intervals to learn a model for automatic speech denoising given only mono-channel audio. Detected silent intervals over time expose not just pure noise but its time varying features, allowing the model to learn noise dynamics and suppress it from the speech signal. An overview of our audio denoise network is shown here:

Our model has three components: (a) one that detects silent intervals over time, and outputs a noise profile observed from detected silent intervals; (b) another that estimates the full noise profile, and (c) yet another that cleans up the input signal.

Dependencies

• Python 3
• PyTorch 1.3.0

You can install the requirements either to your virtual environment or the system via pip with:

pip install -r requirements.txt

Data

Training and Testing

Our model is trained on publicly available audio datasets. We obtain clean speech signals using AVSPEECH, from which we randomly choose 2448 videos (4:5 hours of total length) and extract their speech audio channels. Among them, we use 2214 videos for training and 234 videos for testing, so the training and testing speeches are fully separate.

We use two datasets, DEMAND and Google’s AudioSet, as background noise. Both consist of environmental noise, transportation noise, music, and many other types of noises. DEMAND has been widely used in previous denoising works. Yet AudioSet is much larger and more diverse than DEMAND, thus more challenging when used as noise.

Due to the linearity of acoustic wave propagation, we can superimpose clean speech signals with noise to synthesize noisy input signals. When synthesizing a noisy input signal, we randomly choose a signal-to-noise ratio (SNR) from seven discrete values: -10dB, -7dB, -3dB, 0dB, 3dB, 7dB, and 10dB; and by mixing the foreground speech with properly scaled noise, we produce a noisy signal with the chosen SNR. For example, a -10dB SNR means that the power of noise is ten times the speech. The SNR range in our evaluations (i.e., [-10dB, 10dB]) is significantly larger than those tested in previous works.

Dataset Structure (For inference)

Please organize the dataset directory as follows:

dataset/
├── audio1.wav
├── audio2.wav
├── audio3.wav
...

Please also provide a csv file including each audio file's file_name (without extension). For example:

audio1
audio2
audio3
...

An example is provided in the data/sounds_of_silence_audioonly_original directory.

Data Preprocessing

To process the dataset, run the script:

python preprocessing/preprocessor_audioonly.py

Note: Please specify dataset's directory, csv file, and output path inside preprocessor_audioonly.py. After running the script, the dataset directory looks like the data/sounds_of_silence_audioonly directory, with a JSON file (sounds_of_silence.json in this example) linking to the directory.

Inference

Pretrained weights

You can download the pretrained weights from authors here.

Step 1

1. Go to model_1_silent_interval_detection directory
2. Choose the audioonly_model
3. Run

   CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0,1 python3 predict.py --ckpt 87 --save_results false --unknown_clean_signal true

4. Run

   python3 create_data_from_pred.py --unknown_clean_signal true

5. Outputs can be found in the model_output directory.

Step 2

1. Go to model_2_audio_denoising directory
2. Choose audio_denoising_model
3. Run

   CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0 python3 predict.py --ckpt 24 --unknown_clean_signal true

4. Outputs can be found in the model_output directory. The denoised result is called denoised_output.wav.

Command Parameters Explanation:

1. --ckpt [number]: Refers to the pretrained model located in each models output directory (model_output/{model_name}/model/ckpt_epoch{number}.pth).
2. --save_results [true|false]: If true, intermediate audio results and waveform figures will be saved. Recommend to leave it off to speed up the inference process.
3. --unknown_clean_signal [true|false]: If running inference on external data (data without known clean signals), please set it to true.

Contact

E-mail: [email protected]

© 2020 The Trustees of Columbia University in the City of New York. This work may be reproduced and distributed for academic non-commercial purposes only without further authorization, but rightsholder otherwise reserves all rights.