Adapting-LEAF

The PCEN Adapting experiments under two kinds of noisy environments for INTERSPEECH 2023 Paper.

Paper title: What is Learnt by the LEArnable Front-end (LEAF)? Adapting Per-Channel Energy Normalisation (PCEN) to Noisy Conditions

Authors: Hanyu Meng, Dr Vidhyasaharan Sethu, Prof Eliathamby Ambikairajah

Insititution: The University of New South Wales

Contact: hanyu.meng@unsw.edu.au

Introduction

This is the source code for adapting PCEN layer in LEAF under noisy environment.

Experimental Setups and Results

1. Four model to be trained

Clean Trained: Trained on the entire noise-free training set (baseline).
Noisy Trained: Trained on the noisy version of the entire training data.
Before Adapt: Trained on the noise-free training set without including adaptation data.
PCEN Adapt: The BA model with the PCEN layer was adapted using the noisy adaptation data.

As for the backend, we use the same back-end (EfficientNetB0) as the original LEAF paper.

2. Speech Processing Task: Emotion Recognition

3. Dataset and Partition

We applied the CREMA-D Dataset with the following partition:

4. Noise Setups

We varying the Signal-to-Noise Ratio (SNR): from 0 to 20 dB with 5dB increment (We trained all 4 models under 0 dB, 5 dB, 10 dB, 15 dB, 20 dB SNR correspondingly).
Two representative classes of noise were chosed
- Gaussian Noise (Stationary Noise)
- Babble Noise (Non-stationary Noise)
  - We create the babble noise environment by randomly select 3 speech samples from the MUSAN dataset and mixed them up.
  - The availablity of MUSAN dataset: Download

5. Results and Analysis

We evaluate the accuracy of the overall model accuracy under noisy environments (adding different kinds and level of noise to the test set). The accuracy of all 4 models under different kinds/levels or noise is shown as follow:

Analysis:

Gaussian Noise Adaptation
- Noisy training data helps the model learn the pattern of noise and improves its robustness.
- Adapting the PCEN layer with a small amount of noisy data, the impact of noise on accuracy can be mitigated.
Babble Noise Adaptation
- Training with noisy data is not as effective under babble noise conditions.
  - The reason might due to that compared with gaussian noise, babble noise has greater similarity between noise and speech.
- Adapting PCEN with babble noise might be effective in allowing the model to be used under non-stationary noisy conditions.

Getting Started

Prequest

- python=3.8 
- setuptools==59.5.0
- numpy==1.23.3
- tqdm
- tensorboard
- pytorch==1.10.0
- tensorflow-datasets
- efficientnet_pytorch
- PySoundFile
- soundfile

Installation

git clone https://github.com/Hanyu-Meng/Adapting-LEAF.git

Configrations

Creating Noisy Dataset
- Create Noisy Dataset with Different Level of Gaussian Noise
1. Change the dataset directory and new path in create_noisy_dataset.py to your dataset directory and the directory to store the noisy dataset.
2. Open dataset_pre_process.py, change exec
```
exec = "/your_environment_directory /current_path/create_noisy_dataset_dataset.py"
```
1. Run dataset_pre_process.py
```
python3 dataset_pre_process.py
```
- Create Noisy Dataset with Different Level of Babble Noise
1. Change the dataset directory, new path, and dir in create_babble_noise_dataset.py to your dataset directory, the directory to store the noisy dataset, and your directory for MUSAN speech subset.
```
directory = "/your_path_for_CREMA_D"
path_new = "/your_path_for_different_level_of_Babble_CREMA_D"
dir = "/your_path_for_MUSAN_SPEECH"
```
1. Run dataset_pre_process.py
```
python3 dataset_pre_process.py
```

Baseline Models Training

Clean Trained

Gaussian Noise

python3 main.py --data-set 'CREMAD_SEN_90' --noise_test True  ---babble_test False --noise False --level {noise_level}

Babble Noise

python3 main.py --data-set 'CREMAD_SEN_90'  --noise_test False --babble_test True --babble False --level {noise_level}

Noisy Trained

Gaussian Noise

python3 main.py --data-set 'CREMAD_SEN_90' --noise_test True  ---babble_test False --noise True --level {noise_level}

Babble Noise

python3 main.py --data-set 'CREMAD_SEN_90'  --noise_test False --babble_test True --babble True --level {noise_level}

Alternatively, you can modified the batch-job script batch_job_noisy_train.py to run the experiments, it can automatically vary the noise level and store the models.

Adaption Models Training

Before Adapt (BA)

Gaussian Noise

python3 main.py --data-set 'CREMAD' --noise_test True  ---babble_test False --noise True --level {noise_level} --tune False

Babble Noise

python3 main.py --data-set 'CREMAD'  --noise_test False --babble_test True --babble True  --level {noise_level} --tune False

PCEN Adapt (PA)

Gaussian Noise

python3 main.py --data-set 'CREMAD' --resume {path_to_corresponding_BA_model} --noise_test True  ---babble_test False --noise True --level {noise_level} --tune True

Babble Noise

python3 main.py --data-set 'CREMAD' --resume {path_to_corresponding_BA_model}  --noise_test False --babble_test True --babble True --level {noise_level} --tune True

There are also have batch_job files batch_job_before_adapt.py and batch_job_pcen_adapt.py for reference, you can modify the job scripts according to your settings.

File Structures

├── Adapting-LEAF
│   ├── Job_script
│   │   ├── batch_job.py
│   │   ├── batch_job_before_adapt.py
│   ├── Noise_dataset_create
│   │   ├── image1.png
│   │   ├── image2.jpg
│   ├── Pre-trained models
│   │   ├── Babble Noise
│   │       ├── babble_BA_0db
│   │           ├── args.txt
│   │           ├── final_metrics.txt
│   │           ├── net_checkpoint.pth
│   │           ├── net_last_model.pth
│   │       ├── .....
│   │ 
│   │   ├── Gaussian Noise
│   │       ├── BA_0db
│   │           ├── args.txt
│   │           ├── final_metrics.txt
│   │           ├── net_checkpoint.pth
│   │           ├── net_last_model.pth
│   │       ├── .....
│   │   
│   └── datasets
│       ├── __init__.py
│       ├── crema_d.py
│   └── model
│       ├── __init__.py
│       ├── leaf.py
│   └── result_analysis
│       ├── ....
│   └── engine.py
│   └── main.py
│   └── utils.py

Pretained Models

We also provided the pre-trained models in the folder PCEN_Adapting/Pre-trained models. Inside the folder, the Gaussian Noise folder contains noisy trained models, Before Adapt (BA) models, and PCEN adapt (PA) models for different SNR settings. Those same models for babble noise adapting experiment are in the Babble Noise folder.

Credits

As for the LEAF training, we adapted the code from EfficientLEAF Work.

We made some modifications based on this repository.