Dereverberation MetricGAN-U

Experiment

Step0: Pre-Work

In the training phase of Dereverberation, Pairs of Clean audio stream and Reverberated audio stream need to be prepare in advance.

The audio streams with the same basename are viewed as a pair in the training phase

/path/to/ane/aaa.wav
/path/to/ane/bbb.wav

/path/to/rev/aaa.wav
/path/to/rev/bbb.wav

# in train_list
aaa.wav
bbb.wav

Step1: Data Generator

Data Path

The data is stored in /media/ponddy/DATA2/ponddy_dereverb/weak_reverb , and dereverb/dataset/ponddy_dereverb/mix. You can skip this section by directly using these dataset.

Implementation

It is expensive to collect the reverberated and clean wild data at the same time. The most common way is to synthesize the reverb audio streams.

cd reverb_generator

The dataset rir_noise is used for reverberation generation, the dataset is stored at /media/ponddy/DATA2/dereverb/Dataset/RIRS_NOISES and dereverb/dataset/RIRS_NOISES or it can be directly downloaded by the link.

wget http://www.openslr.org/resources/28/rirs_noises.zip

Then set the path in the gernerator.py to the corresponding one.

vim reverb_generator/generator.py

# change the path to the correct path if not running in the provided system.
ps_path = 
rir_path =

The script generate point source reverberation and RIR into a clean audio stream.

Usage:

python main.py --input_path /media/DATA/dayiData/VCTK/wavs \
               --output_path /media/ponddy/DATA2/ponddy_dereverb \
               --mode split \
               --generate \
               --mklist
'''
input_path:  the path of the clean audio files 
output_path: the path to output reverb and clean wav
/path/of/the/output_path　─── rev # reverb 16k Hz wave
                          ├── ane # clean 16k Hz wave

mode:        use data of train, dev, or test. set split can automatically
             seperate the dataset to these three parts
generate:    set this flag to generate the reverb data
mklist:      output a list of training and develop audio files (may used in
             some training)
'''

Other dataset

VoiceBank Datset can be downloaded from http://140.109.21.234:5000/fsdownload/xuC7hkiDg/reverb-vctk-16k. It is stored in dereverb/dataset/reverb-vctk-16k It will be used for the training of speechbrain.

Step2 Train (MetricGAN-U)

Code preparation

The implementation of MetricGAN-U takes advantage of the open-sorce project speechbrain.

To start with Cloning the script from speechbrain.

git clone https://github.com/speechbrain/speechbrain.git

Train

• Find the target repo

cd dereverb/speechbrain/recipes/Voicebank/dereverb/MetricGAN-U

Train with voicebank dataset

python train.py hparams/train_dereverb.yaml --data_folder dereverb/dataset/reverb-vctk-16k

Train with ponddy data

# ponddy_reverb_prepare.py 和 README.md 同目錄，複製到 speechbrain/recipes/Voicebank/dereverb/MetricGAN-U 底下
cp tool/ponddy_reverb_prepare.py /path/to/speechbrain/recipes/Voicebank/dereverb/MetricGAN-U/ponddy_reverb_prepare.py

cd /path/to/speechbrain/recipes/Voicebank/dereverb/MetricGAN-U/ 
vim train.py 
# from voicebank_revb_prepare import prepare_voicebank  # noqa (line 719)
from ponddy_reverb_prepare import prepare_voicebank  # noqa

• Train with ponddy dataset

python train.py hparams/train_dereverb.yaml --data_folder dereverb/dataset/ponddy_dereverb/mix

• inference (evaluation)

  • Test phase follow the completion of the training phase. Commenting the training code leads to access the evaluation part directly.
  • 將train.py 裡面的 se_brain.fit(...) 註解掉，即可跑驗證

• inference (單筆音檔 evaluation)

  • 訓練完成後，會在 speechbrain/recipes/Voicebank/dereverb/spectral_mask/results/spectral_mask/<訓練設定的seed>/save 資料夾底下，找到你的模型 checkpoint 資料夾。

  • 將checkpoint資料夾內的 generator.ckpt 重新命名為 enhance_model.ckpt。

  • 接著請參考 /home/vincent0730/chinese-voice-scoring-widget/English/metricgan_inference.py，其中7-10行設定成你的 enhance_model.ckpt 保存位置，資料夾內還須附上 hyperparams.yaml，hyperparams.yaml 直接沿用即可。

Result

• MetricGanU pesq = 2.03 for Voicebank, pesq = 1.74 for ponddy dataset.
• spectral-mask pesq = 2.35 for Voicebank, pesq = 2.03 for ponddy dataset.

Deployment

Gunicorn

git clone https://github.com/ponddy-edu/ML_web_dereverb.git
cd ML_web_dereverb/
python3.8 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
gunicorn app:app 或是 python app.py

# Reloads the configuration, starts the new worker processes, and gracefully shutdowns older workers
# 通常使用這個指令重啟即可
kill -SIGHUP $(cat gunicorn_pid)

# 強制關閉所有名為gunicorn的服務 (可能會關閉其他的gunicorn，謹慎使用)
pkill -f gunicorn 

• Open your browser and navigate to http://192.168.1.27:8864

Docker Deployment

sudo docker build --no-cache -t "ponddy/dereverb" -f Dockerfile .
sudo docker run -it -d --rm --name ponddy_dereverb_web_flask -p 8864:8864 ponddy/dereverb:latest
sudo docker tag image_id ponddy/dereverb:v1.2
sudo docker push ponddy/dereverb:v1.2
sudo docker push ponddy/dereverb:latest

• Open your browser and navigate to http://192.168.1.27:8864

Report

GOP score table

• Pick 5 users' records
• Best score is marked in bold

	Original	Loud Norm	MetricGAN (sup)	Reverb CLS + MetricGAN (sup)	MetricGAN (unsup)	MetricGAN (unsup) + Synthesized reverb
4405語者	79	90	88	80	90	84
4420語者	73	85	79	71	86	80
4451語者	62	72	68	62	72	73
4479語者	70	79	76	69	86	78
4516語者	69	77	74	78	82	76

Kullback–Leibler(KL) divergence table

• A measure of how one probability distribution Q is different from a second, reference probability distribution P.
• KL divergence 用來量測兩個分布之間的距離，0為最小值，表示兩分布相同
• Reverb CLS + MetricGAN (sup) 和 Original 的分布相差最遠，同時 Reverb CLS + MetricGAN (sup) 的分布更靠左一些

	Loud Norm	MetricGAN (sup)	Reverb CLS + MetricGAN (sup)	MetricGAN (unsup)	MetricGAN (unsup) + Synthesized reverb
Original	0.00206	0.00383	0.01041	0.00328	0.0030

分布圖說明

• 圖內有兩個分布重和，分別是Original(偏橘色)，以及測試的方法(藍色)，而重疊部分為紫色
• 當測試的方法(藍色)在分布右邊(高分)出現越多，代表分數的分布提升
• 當測試的方法(藍色)在分布左邊(低分)出現越多，代表分數的分布降低

Loudness normalization

• Thoughts
  • Can be easily integrated into our current system
  • Lowest latency compared to other deep learning solution
  • GOP 訓練資料並無進行 loudness normalization 預處理，可列入考慮

MetricGAN (supervised) - 人工殘響進行訓練

• Thoughts
  • 有些許成效，但訓練資料不易取得，需準備大量乾淨音檔(困難)以及近似使用者的迴響情境(困難)

Reverb classification & MetricGAN (supervised) - 人工殘響進行訓練

• Thoughts
  • 效果不佳，多數使用者音檔均被歸類為有迴響情境

MetricGAN (unsupervised) - 真實5000筆殘響進行訓練

• Thoughts
  • 真實使用者音檔資料易取得，不需要人工合成殘響即可訓練
  • 長遠使用來看，使用非監督式模型效果較佳

MetricGAN (unsupervised) - 真實5000筆殘響 + 人工殘響進行訓練

• Thoughts
  • 效果和單純使用 MetricGAN (unsupervised) 差不多，但加入多種殘響能增加泛化能力