[From Bruce Nevin (980507. EDT)]
Speakers produce distinctions between words without being aware that they are making them. I understand that many examples are detailed in Bill Labov's book Principles of Language Change, but I have only seen a couple of early papers on it. In all cases there are bona fine minimal pairs, that is, different words that differ only in one consonant or vowel segment. In an English example, the [aj] - [oj] diphthongs undergo near merger and there are mininal pairs like <line> vs. <loin>, which the relevant speakers pronounce differently but are unable to discriminate. In Philadelphia there is a similar situation with <ferry> vs. <furry>. In some parts of the American west it is <pull> vs. <pool>, etc.
Speakers with near mergers are always found in communities where some speakers have a complete merger and some others make a clear distinction.
Here is what I think is going on. The speakers in Philadelphia, for instance, learned to pronounce <ferry> and <furry> differently, but most people around them didn't make any distinction between these two vowels. The reference values for auditory perceptions are re-set, but the recalibration of reference levels for articulatory perceptions (how much pressure sensation at what points, etc.) leaves traces of an earlier distinction whose acoustic effects are no longer audible. The next generation of children sets articulatory reference levels in order to make the distinctions that they hear made by their elders, so the "ghost" articulation disappears.
If so, then this phenomenon is not found in children's speech. I will send mail to Bill Labov and see if he can verify this.
Bruce Nevin+