Joint training embeddings

---

This is the official implementation of the submitted Interspeech 2022 paper [Enhancing Embeddings for Speech Classification in Noisy Conditions], (Mohamed Nabih Ali, Daniele Falavigna, Alessio Brutti).

Description

---

In this paper, we investigate how enhancement can be applied in neural speech classification architectures employing pre-trained speech embeddings. We investigate two approaches: one applies time-domain enhancement prior to extracting the embeddings; the other employs a convolutional neural network to map the noisy embeddings to the corresponding clean ones. All the experiments are conducted based on Fluent Speech commands, Google Speech commands v 0.01 and generated noisy versions of these datasets.

Architectures

Dataset

You need to create .txt files for training, validation, and testing datasets with the following structure:

<noisy_1_path><space><clean_1_path> 
<noisy_2_path><space><clean_2_path> 
<noisy_n_path><space><clean_n_path>

e.g. In the case of Wave-Enh strategy

/train/noisy/a.wav /train/clean/a.wav
/train/noisy/b.wav /train/clean/b.wav

In the case of the Embed-Enha strategy, it should be like 

/train/noisy/a.pt /train/clean/a.pt
/train/noisy/b.pt /train/clean/b.pt

For the Wave-Enh the test.txt file should be

<noisy_1_path>
<noisy_2_path>
<noisy_n_path>
e.g.
/test/noisy/a.wav
/test/noisy/b.wav

Steps of Wave-Enh Strategy

You need to download the wav2vec model from (https://github.com/pytorch/fairseq/tree/main/examples/wav2vec), and modify its path in util/utils.py file.

Training

Use train.py to jointly train both the speech enhancement and the speech classifier modules. It receives two command line parameters:

• -C, --config, the path of your configuration file for the training process.
• -R, --resume, resume training from the last saved checkpoint.

Syntax python train.py -C config/train/train.json or python train.py config/train/train.json -R

Evaluation

Use enhancement.py to evaluate both models.

• -O specify the folder where to save the enhanced signals.
• -D use the signals using the GPU.
• -M path to save the best front-end model.
• -m path to save the best back-end model.

Syntax: python enhancement.py -C config/enhancement/unet_basic.json -D 0 -O <path to save the enhanced signals> -M <path to the best speech enhancement model> -m <path to the best back end speech classifier>

Steps of Embeds-Enh Strategy

Embeddings Extraction

---

You need to download the wav2vec model from (https://github.com/pytorch/fairseq/tree/main/examples/wav2vec), and modify its path in emb.py file to extract the embeddings from the dataset.

Syntax: python emb.py

Training

Use train.py to jointly train both the speech enhancement and the speech classifier modules. It receives six main commands line parameters:

• -m path to save the best back-end model.
• -M path to save the best front-end model.
• -b number of residual blocks in the back-end speech classifier.
• -r number of repeats of the residual blocks.
• -lr learning rate
• e number of epochs

Syntax: python train.py -m ./best_bkmodel.pkl -M ./best_frmodel.pkl -b 5 -r 2 -lr 0.001 -e 100

Evaluation

Use evaluation.py to evaluate both models based on the test dataset. It receives two main commands line parameters:

• -m path to save the best back-end model.
• -M path to save the best front-end model.

Syntax: python evaluation.py -m ./best_bkmodel.pkl -M ./best_frmodel.pkl