Multilingual Speech Language Identification

Machine Learning Final Project

Overview

This project implements an end-to-end spoken language identification system using classical machine learning techniques. It is developed in two phases:

• Phase 1: Dataset Collection and Problem Understanding
• Phase 2: Processing, Modeling, and Evaluation

The goal is to classify short speech segments into one of four languages and analyze their structure using both supervised and unsupervised learning.

---

Phase 1 – Dataset Construction

A multilingual speech dataset was created from audiobook-style podcast recordings in:

• Italian
• German
• Korean
• Spanish

Each audio file:

• Is approximately one minute long
• Starts at the beginning of a sentence
• Ends at the end of a sentence
• Contains clean, continuous speech

This phase focused on building a balanced and structured dataset suitable for machine learning.

---

Phase 2 – Machine Learning Pipeline

In this phase, the collected audio data was processed and analyzed through:

• Data cleaning and preprocessing
• Feature extraction
• Supervised classification (multiple ML models)
• Unsupervised clustering
• Quantitative evaluation (Accuracy, F1-score, Confusion Matrix, Silhouette Score)

The project demonstrates a complete workflow from raw speech data to language classification and analysis.