README.md

Web speech denoiser and transcriber

python 3.9

Структура проекта:

.
├── app # web and live applications
├── doc # some samples, license, presentation
└── wave_wizard # modeling, fitting, test and so on.

Результаты обучения

Запуск приложений

Инструкция по запуску:

git clone https://github.com/sir-timio/web-denoiser.git
cd web-denoiser
pip install -r requirements.txt 

download/get clean and noise audio files. Used data: https://github.com/microsoft/MS-SNSD:

    git clone https://github.com/microsoft/MS-SNSD
    mv MS-SNSD/noise_train wave_wizard/noise_train
    mv MS-SNSD/noise_train wave_wizard/clean_train

optional: add to clean_data some subsets from https://github.com/snakers4/open_stt/#links with 99%+ quality and to noise_data files from https://www.kaggle.com/datasets/chrisfilo/urbansound8k

generate train data, split it into train, val and test as you wish, but test metrics take a lot of time

    cd wave_wizard
    python noiser.py --config configs/noiser_config.yaml

basic training: run, check metrics and debug samples

    python train.py --config configs/train_config.yaml

advanced training with clearml queue and triggers:

run first experiment to get ID for src/triggers/dataset_trigger.py

    python train.py --config configs/train_config.yaml

add dataset with clearml-data, example:

    clearml-data create --project<project_name> --name <dataset_name>
    clearml-data add --files <path_to_data>

configure src/triggers/dataset_trigger.py

clearml-data sync --project <project_name> --name <dataset_name> --parent <parent_dataset_id> --folder <path_to_new_data>

I recommend name child dataset as parent, thus you could get newest version of dataset by name.

and finally, run task, manage queue with web.

    python src/triggers/dataset_trigger.py

As soon as dataset would be updated, task would automatically started again.