id:	3
title:	HVPT phonetic word pair training

Note: Initial project design page here.

Original Language Log post which prompted me to start thinking of this - it mentions Dave Pisoni!

Turns out people can learn to distinguish foreign phonemes pretty easily when they hear lots of pronunciations in sequence and have to choose which one they're actually hearing. This is HVPT (High Variability Phonetic Training.) A system to train these would:

1. Allow a user to practice anonymously, or with an identity. If with an identity, the system could maintain a session, and provide reporting after conclusion of the session. The report could be emailed directly to a language teacher for verification; this allows the system's use in homework.

2. For each phoneme pair of interest, the system would have a series of minimally differing pairs of words (e.g. "mitt" vs. "meet" or "sheep" vs. "ship".) For each word, the system has a number (at least 20) examples of different people pronouncing the word. There would have to be a distributed way to collect those audio snippets. (Simple upload is fine, of course.)

3. A practice round simply consists of the random selection of a word pair for a phoneme pair in question, and the system plays an audio pronunciation. (See "shiporsheep.com" for technical example.) The user clicks the word that corresponds to the audio heard. If the user's right, immediate payoff; if not, immediate correction (by pronouncing the word clicked -- if possible, pronounced by the same speaker.)

4. A session can mix and match phoneme pairs; sets of troublesome phoneme pairs should be grouped by the student's language, or simply by student selection. Scores could also influence the selection of troublesome phonemes (harder ones coming up more often.)

5. You could easily make a CD with one track for each phoneme pair, simply consisting of a sequence of audio snippets. Download the MP3s and burn your own CD (then copy to tape.)

... I think that's it.