Unsupervised Lexicon Learning in Continuous Speech

Title: Unsupervised Lexicon Learning in Continuous Speech

Abstract: This study investigates an unsupervised approach to discovering words in continuous spoken speech. The algorithm is based on a traditional statistical model and competes with other recent algorithms that require supervised learning. The main contributions of this research are:

1. Demonstrating the applicability and competitiveness of a traditional approach for a task that has been approached nontraditionally. 2. Showing that the algorithm's performance remains consistent with results from learning theory, even with partial supervision.

The algorithm is incremental, meaning it processes the speech in small chunks and learns from each piece, making it truly unsupervised. It does not require tunable parameters or multiple passes over the data, making it more efficient and flexible than other unsupervised methods.

Link to Article: https://arxiv.org/abs/0111064v1 Authors: arXiv ID: 0111064v1