rg_speech_to_text

Research Group Speech To Text

This repository will be used to experiment/test new approaches before they are fit into the common codebase.

Environment preparation

This repo can be used in native Windows 10, native Ubuntu, Mac. Some features related to language models requires Linux. For Windows 10 users there is a possibility to use Ubuntu under WSL2.

Windows 10

Installing via pip

• Download and Install python (recommend 3.8)
• Create a virtual environment using python -m venv env_name
• enable created environment env_path\Scripts\activate
• Update pip and setuptools using python -m pip install --upgrade pip setuptools
• Install PyTorch 1.7.1 from https://pytorch.org/get-started/locally/ with torchvision and torchaudio pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html
• install required dependencies pip install -r requirements-win.txt

Installing via conda

• Download and install miniconda
• Create a new virutal environment using conda create --name env_name python==3.8
• enable create environment conda activate env_name
• Install PyTorch conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch
• install required dependencies pip install -r requirements-win.txt For Windows platform it is enough requirements.txt and in addition install the audio driver from conda, like the following:

  conda install -c conda-forge python-sounddevice
  conda install pyaudio
  

  The reason for this different install path is explained here.

[optional] WSL2 setup

If not using WSL2, skip this section.

For users of WSL2, we provide a helper setup script available in wsl2_setup directory. That scripts will create a dedicated wsl instance, which then can be used for this project. Depending on which Windows version we have, we distinguish the following:

• Windows Build >= 20150 have WSL2 with GPU access, therefore it can be installed CUDA as in native Linux.
• Windows Build >= 21376 have WSL2 with WSLg, PulseAudio server integrated to communicate with host OS audio and can run graphical Linux apps.
• Windows Build < 20150 have WSL2 which is sufficient to run in CPU mode. With minimal effort, WSL2 of any version listed above can give us an Ubuntu 20.04 LTS environment in which the OS specific setup is exactly the same as in native Ubuntu 20.04 LTS. Therefore, the next section is also necessary to be executed in the running WSL2 instance.

[optional] Ubuntu 20.04 LTS setup

If not using Ubuntu (including Ubuntu in WSL2), skip this section.

The setup script below is common for native Ubuntu or Ubuntu running in WSL2.

ubuntu_setup/initial_setup.sh will

• install the necessary ubuntu packages,
• create the python virtual environment,
• install the requirements.
• install NeMo toolkit r1.0.0rc1 from github

Installing TheSoundOfAIOSR's rg_speech_to_text from source

bash
git clone https://github.com/TheSoundOfAIOSR/rg_speech_to_text
cd rg_speech_to_text
pip3 install .