Proceedings Fonetik 2009 - Institutionen för lingvistik

Proceedings, FONETIK 2009, Dept. of Linguistics, Stockholm Universityeither as short or long, e.g. baka was pronouncedas either [baaka] (eight speakers) or[baakka] (four speakers). However, this alternationhad no effect on VOT].) Word-final /ptk/were fully voiceless. To the ears of the secondauthor, the Fenno-Swedish /p t k/ sound verymuch like Finnish /p t k/.The set /b d g/In the realization of the utterance-initial /b d g/,there was much variation among the speakers.For eight of the twelve speakers, 95.2% of theutterance initial lenis tokens were prevoiced,whereas for the remaining four speakers, only70.6% of the tokens were prevoiced. Half of thespeakers in both subgroups were female andhalf were male. For the group of eight speakers,the mean VOT was -85 ms, and the proportionof tokens with non-negative VOT was 4.8%.The results for this group are similar to those ofHelgason and Ringen (2008) for CS Swedishwho observed the grand mean VOT of -88 msand report that, for all six speakers pooled, 93%of the initial /b d g/ had more than 10 ms prevoicing.For the CS Swedish speaker with theshortest mean prevoicing, 31 of her 38 initial /bd g/ tokens had more than 10 ms of prevoicing.For our group of eight Fenno-Swedish speakersand for all of the six CS Swedish speakers inHelgason and Ringen, then, extensive prevoicingwas the norm, with only occasional nonprevoicedrenditions.For the group of four Fenno-Swedishspeakers the mean VOT was only -40 ms, andthe proportion of tokens with non-negativeVOT was 29.4%. For an extreme speaker inthis respect, the mean VOT was -28 ms and38% of the /b d g/ tokens had non-negativeVOT. In fact, at least as far as VOT is concerned,many /b d g/ tokens overlapped phoneticallywith tokens of /p t k/. A linear discriminantanalysis was run on all initial stopsproduced by the group of eight speakers and onall initial stops produced by the group of fourspeakers to determine how well the analysiscan classify the stop tokens as instances of /p tk/ or /b d g/ on the basis of VOT. For the groupof eight speakers, 97.3% of the tokens werecorrectly classified: the formally /p t k/ stopswere all correctly classified as /p t k/, 4.8% ofthe formally /b d g/ stops were incorrectlyclassified as /p t k/. For the group of fourspeakers, 82.9% of the tokens were correctlyclassified: 1.0% of the formally /p t k/ stopswere incorrectly classified as /b d g/, and 29.4%of the formally /b d g/ stops were incorrectlyclassified as /p t k/. For these four speakers,then, /b d g/ often had positive VOT values inthe small positive lag region also frequently observedin /p t k/.Medial /b d g/ were extensively or fullyvoiced. In 9.4% of the tokens, the voiced proportionof the occlusion was less than 50%, in10.4% of the tokens the voiced proportion was50% or more but less than 75%, in 10.4% of thetokens the voiced proportion was 75% or morebut less than 100%, and in 70.0% of the tokensthe voiced proportion was 100%. Thus, whilethe medial /b d g/ were on the whole veryhomogeneous and mostly fully voiced, 3.6% ofthem were fully voiceless. These were nearlyall produced by three of the four speakers whoalso produced many utterance-initial /b d g/ tokenswith non-negative VOT values. The fullyvoiceless tokens were short and long /d/’s andshort /g/’s; there was no instance of voiceless/b/. A discriminant analysis was run on all medialstops, with closure duration, speaker sex,quantity, place of articulation, duration of voicingduring occlusion and positive VOT as theindependent variables. 96.4% of the tokenswere correctly classified (98.7% of /p t k/ and94.5% of /b d g/). The order of magnitude ofthe independent variables as separators of thetwo categories was duration of voicing duringocclusion > positive VOT > closure duration >place > quantity: sex had no separating power.Great variation was observed in both shortand long final /b d g/, and therefore the speakerswere divided into two subgroups both consistingof six speakers. This was still somewhatProcrustean, but less so than making no divisionwould have been. For group A the meanvoiced proportion of the occlusion was 89%(s.d. = 18%), for group B it was 54% (s.d. =31%). As the standard deviations suggest, therewas less inter-speaker and intra-speaker variationin the A group than in the B group.Among the group A speakers, mean voicedproportion of occlusion across the places of articulationranged from 73% to 98% in the short/b d g/ and from 63% to 99% in the long /b d g/,among the group B speakers the correspondingranges were 36% to 62% and 46% to 64%. Anextreme example of intra-speaker variation is amale speaker in group B for whom four of the24 final /b d g/ tokens were completely voicelessand nine were completely voiced. Discriminantanalyses were again run on all finalstops, separately for the two groups. For groupB, with voicing duration during the occlusion,61