A method for aligning a transcript to recorded speech is described, motivated by the need for alignment of audiobooks for use by SLA software. A combination of text-to-speech technology, mel-spectrum feature extraction, and dynamic time warping are employed to obtain a word-alignment for the input speech sample.
I'm curious about unsupervised word sense disambiguation, and unsupervised machine learning in general. For that, Manning and Schuetze tell us we need clustering. I jumped ahead to Chapter 14 to experiment with clustering algorithms.
The other day I picked up my Chinese copy of Alice in Wonderland that I picked up in Beijing last year. My intention was to lay in the sun by the lake until I had finished the first page, using the dictionary as needed to achieve basic comprehension. The result was a bad sunburn and only two of four paragraphs finished. What went wrong?
Over the past few days I've worked through Chapter 2 of Foundations (Manning and Schuetze), including writing a Python program to reproduce Shannon's Experiment.
First an explication: a few weeks ago I had a revelation that computational linguistics is the field for me. So, I found the best book I could on the subject, which seems to be Manning and Schuetze's Foundations of Statistical Natural Language Processing (FSNLP). Just reading the introduction on Google Books was enough to convince me to drop some hard cash. The writing is brilliant. I'm only through Chapter One, but it's already worth it.