jinhxu / speaker-identification-system-prototype

A high-performance (98% identification accuracy) & low-cost industrial speaker identification system.

Speaker Identification (98% identification accuracy)

How to build a high-performance & low-cost industrial (data-driven) speaker identification system with 98% identification accuracy using the state-of-the-art algorithms implemented with open-source packages and frameworks? This thesis describes an extensive research conducted towards the goal of building such a system prototype. The research answers the following questions:

• what are the state-of-the-art algorithms? NN-based SpeakerNet vs. classic ML MFCC-GMM
• minimum training data? how does the model performance change over different amount of training data (closed-set vs. open-set)
• minimum inference data? how does the model performance change over different amount of inference data (closed-set vs. open-set)
• in an open-set identification enviroment, instead of setting a confidence score threshold, how does using a class of various speaker recordings to represent un-enrolled users?

prototype system specifics

• minimum required training data per user: 3 min
• minimum required data for each inference on the run: 3 seconds
• system: ROS (Robot Operating System)

Abstract

Speaker identification has numerous applications in various areas. Over the years, there has been a rising need for automated speaker identification systems in the industry. This thesis explores text-independent speaker identification on short audio samples using data-driven methods. In particular, it concentrates on the available solutions with high identification accuracy and seeks to build a high-performance practical system. The experiments compare two of the state- of-the-art statistical models for the task, examine the effect of the amount of training data on model performance in closed-set and open-set settings, and ex- amine the effect of marginal change on inference audio length on identification accuracy. The highest accuracy achieved by both models in the experiments surpass 0.985 in a closed-set setting, and 0.978 in an open-set setting. Beyond discussing the observations and findings from the experiments, this thesis describes the practical speaker identification systems built in the process based on the experiment results and provides a reproducible pipeline for building a high-performance practical speaker identification system.

View the Thesis

Author: Jinghua Xu

Links

• Tutorial: speaker identification using SpeakerNet

• Convert Large MD tables to Latex (automatic parser)

This thesis received a grade of 1.0/1.0 (German scale) at the University of Tübingen.