Speech analysis pipeline

Preparing the environment

1. Download the code

Clone this git repo locally.

$ git clone https://github.com/rudn-speech-analysis/analysis-node
$ cd analysis-node

Update the submodules.

git submodule init && git submodule update --recursive

2. Create the conda environment

$ conda env create -f ./environment.yaml

Activate the environment:

$ conda activate speech-analysis

3. Install flash-attn

Nvidia

For Nvidia GPUs installing directly with pip or conda is enough

AMD

For AMD GPUs build flash-attn yourself.

speech-analysis environment already defines some flash-attn's environment variables, but you can set them yourself:

$ export FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE
$ export FLASH_ATTENTION_TRITON_AMD_AUTOTUNE=TRUE

Then build the library:

$ cd flash-attention
$ python setup.py install