fairseqAPI

Small example use of the fairseq python library

This is a proof of concept of how the fairseq python library can be used for tts, given that the examples on HuggingFace and the repo are either unclear, or broken.

Installation

(This installation should work for python versions 3.8 through 3.10. version 3.11 is not yet supported. Run python --version to view current installation)

clone this repository to your local machine and cd into it

git clone https://github.com/Alex23rodriguez/fairseqAPI.git
cd fairseqAPI

clone facebook's fairseq repository, as there are many modules that we use. Notice that we change the name to avoid confision between the module and the repo folder.

git clone --depth 1 https://github.com/pytorch/fairseq fairseq_repo

NOTE: this is different from pip install fairseq, which installs the CLI tool, not the python module. Check out their repo for more info.

install the required modules
- it is recommended to create a virtual envirenment before installing requirements: python -m venv venv && source venv/bin/activate

pip install -r requirements.txt

Usage

downloading models

Once the setup is done, we can move on to downloading a model from Hugging Face. Given that I personally use this for Russian, the model defaults to this one (note that the example on HG's website is currently broken)

to download the russian tts model, simply run the following command:

python download_hf_model.py

to download any other model, pass it as a commandline argument

python download_hf_model.py facebook/tts_transformer-ru-cv7_css10

note that models are typically heavy, which is why we only want to do this step once

using models

To use tts, simply call tts.py with the corresponding commandline arguments (python tts.py -h for more info).

python tts.py Здравствуйте, это пробный запуск. До свидание.

output defaults to ./out/out.wav
- to play the output from the terminal install sox and use play out/out.wav
- sox can also be used to change the speed of the recording: sox input.wav output.wav tempo 1.33