SpeakerVerSim

Overview

SpeakerVerSim is an easily-extensible Python-based simulation framework for different version control strategies of speaker recognition systems under different network configurations.

These simulations are used in the paper Version Control of Speaker Recognition Systems.

How to use

Install

The SpeakerVerSim library can be installed with:

pip install SpeakerVerSim

Run one simulation

You can easily start a simulation by running:

python run_simulator.py

This simulation will use the configurations in the example_config.yml file.

To use a different configuration file, you can use the -c or --config flag. For example:

python run_simulator.py --config my_config.yml

You can also override the strategy to be simulated from the config file using the -s or --strategy flag. For example:

python run_simulator.py -c my_config.yml -s SSO-sync

Call the API

The highest level API is the SpeakerVerSim.simulate function. It takes either the path to the config file or an object of the configurations as its input, and outputs the simulation stats.

Example usage:

import SpeakerVerSim

stats = SpeakerVerSim.simulate("example_config.yml")

Reproduce experiments

You can easily reproduce the experiments in our paper by running:

python run_exp.py

After running this script, simulation results will be stored in the result_stats directory.

Then you can visualize the metrics by running:

python visualize_results.py

The visualization graphics will be stored in the figures directory.

List of implemented strategies

Design

The design of this library is summarized as below:

• This library is built on top of SimPy, a process-based discrete-event simulation (DES) framework based on standard Python.
• All configurations of the simulation are represented in a single YAML file. example_config.yml has explanations for all the configuration fields.
• Each machine in the network inherits from the Actor class, including the client, the frontend server, the cloud worker, and the database.
• All clients inherit from the BaseClient class; all frontend servers inherit from the BaseFrontend class; all cloud workers inherit from the BaseWorker class; and all databases inherit from the BaseDatabase class.
• The communication between two machines happens like this: the sender creates a Message object, and adds it to the receiver's message pool, which is a simpy.Store object.
• During the simulation, metrics are logged in an object of the GlobalStats class.
• The entire network system is represented by the NetworkSystem class or its subclass.

Each version control strategy is implemented by creating a set of client, frontend server, cloud workers, database, and defining how they interact with each other.

Citation

@article{wang2020version,
  title={Version control of speaker recognition systems},
  author={Wang, Quan and Moreno, Ignacio Lopez},
  journal={arXiv preprint arXiv:2007.12069},
  year={2020}
}