vistec-ser 0.4.0a4

Speech Emotion Recognition models and training using PyTorch

Vistec-AIS Speech Emotion Recognition

Speech Emotion Recognition Model and Inferencing using Pytorch

Installation

From Pypi

pip install vistec-ser

From source

git clone https://github.com/tann9949/vistec-ser.git
cd vistec-ser
python setup.py install

Usage

Training with AIS-SER-TH Dataset

We provide Google Colaboratory example for training the AIS-SER-TH dataset using our repository.

Training using provided scripts

Note that currently, this workflow only supports pre-loaded features. So it might comsume an additional overhead of ~2 Gb or RAM. To run the experiment. Run the following command

Since there are 80 studios recording and 20 zoom recording. We split the dataset into 10-fold, 10 studios each. Then evaluate using k-fold cross validation method. We provide 2 k-fold experiments: including and excluding zoom recording. This can be configured in config file (see examples/aisser.yaml)

python examples/train_fold_aisser.py --config-path <path-to-config> --n-iter <number-of-iterations>  

Inferencing

We also implement a FastAPI backend server as an example of deploying a SER model. To run the server, run

cd examples
uvicorn server:app --reload

You can customize the server by modifying example/thaiser.yaml in inference field.

Once the server spawn, you can do HTTP POST request in form-data format. and JSON will return as the following format:

[
  {
    "name": <request-file-name>,
    "prob": {
      "neutral": <p(neu)>,
      "anger": <p(ang)>,
      "happiness": <p(hap)>,
      "sadness": <p(sad)>
    }
  }, ...
]

See an example below:

Author & Sponsor

Chompakorn Chaksangchaichot

Email: [e-mail]