The introduction, technical details, and results are presented in the wandb report.

To get started install the requirements

pip install -r ./requirements.txt

SpEX+ architecture with additional speaker classification head was implemented in this project.

To train model from scratch run

python3 train.py -c final_model/config.json

For fine-tuning pretrained model from checkpoint, --resume parameter is applied. For example, fine-tuning pretrained model organized as follows

python3 train.py -c final_model/finetune.json -r saved/models/pretrain_final/<run_id>/model_best.pth

This command generates new mixed dataset. This option can be disabled by passing "reuse": true for train dataset in config final_model/finetune.json.

Before applying model pretrained checkpoint is loaded by python code

import gdown
gdown.download("https://drive.google.com/uc?id=19i4NIk8R8AlkGvMfhQl8ex-eCg4g2Isv", "default_test_model/checkpoint.pth")

Model evaluation is executed by command

python test.py \
   -c default_test_model/config.json \
   -r default_test_model/checkpoint.pth \
   -t test_data \
   -o test_result.csv \
   -g <output_dir> \
   -s <interval_len>

Where -o specify output .csv file, which represents metrics

• PESQ (Perceptual Evaluation of Speech Quality)
• SI-SDR (Scale-Invariant Signal-to-Distortion Ratio)

Further sections reveal other command line arguments.

Important remark: for further experiments test-clean part of Librispeech dataset is required. It will be automatically installed after at least one execution python3 test.py with default arguments.

Model evaluation with custom data conducted by running test.py

python3 test.py -t path/to/custom/dir

This command executes model on custom dataset folder, which includes mix, target and ref subdirectories with filenames *-mixed.wav, *-target.wav and *-ref.wav respectively. Such directory will be created in data/datasets/mixed/data/custom from mixed test-clean dataset after running

bash custom_set.sh

Extracted audio for the test set can be gathered in one directory by executing

python3 test.py -g path/to/output/dir

This results can be compared with direct speech recognition from mixed audio. Speech recognition pipeline was taken from asr repository. Comparison of mixed and extracted audios quality carried out by

bash asr_score.sh

For training stability audio were split into 3-seconds interval. However, test data provides arbitrary lengths of audios, which can be divided into intervals on inference stage with

python3 test.pt -s <interval_len_in_seconds>

WHAM! dataset provides diverse background noise, which can be also mixed with input audio. Installation

wget https://my-bucket-a8b4b49c25c811ee9a7e8bba05fa24c7.s3.amazonaws.com/wham_noise.zip
unzip wham_noise.zip

Creating noised dataset and model evaluation

python3 test.py -c wham_test/config.json

Described pipeline only involves tt (test) part of WHAM!, therefore, other directories are not required.

This repository is based on an asr-template repository.