Phonetics

Speech detection, such as those used in speech-to-text systems, is a surprisingly difficult problem. There are so many variations in styles of speaking, pronunciation, dialect, and accent, as well as variations in rhythm, tone, speed, and elocution, plus the fact that audio is a simple one-dimensional time-domain signal, that it's no surprise that even today's state-of-the-art smartphone tech is good, not great.

While modern speech-to-text goes much deeper than what I'll present here, I would like to show you the concept of phonetic algorithms. These algorithms transform a word into something resembling a phonetic hash, such that it is easy to identify words that sound similar to one another.

The metaphone algorithm is one such phonetic algorithm. Its aim is to reduce a word down to a simplified phonetic form, with the ultimate goal of being able to index similar pronunciations. Metaphone uses an alphabet of 16 characters: 0BFHJKLMNPRSTWXY. The 0 character represents the th sound...