Human Judgement

Among automatically-detected pronunciation errors, seven samples for /h/ in [yoNhoN], /k/ in [saNkai], /p/ in [saNpuN] are assured to be errors by all five human raters as shown with a dotted line in Table 3.4. For the remaining errors, their judgements were ambiguous. No mispronunciation was found in ichi+compounds, which is contrary to our expectations. It may be because most of non-native learners memorize well this first-beginning irregularity.

(1)

First, learners A and C unclearly pronounced /h/ in [yoNhoN]. A voiceless sound after /N/ is known to be hard to pronounce (see TYPE 7 in section 3.6.3).

(2)

Learners A, B and D pronounced /g/ for /k/ in [saNkai]. It is assumed that the homonym, 三階 (third floor) that can be read as [saNgai], affected their mispronunciation.

 

SYM NATIVE A B C D E F G H

iqpoN -27.47 -28.14 -26.86 -24.69 -27.91 -26.86 -28.78 -25.16 -27.55

nihoN -26.37 -33.51 -28.89 -24.98 -26.50 -28.89 -26.98 -25.10 -25.94

saNboN -26.48 -28.12 -26.84 -25.35 -27.81 -26.84 -25.83 -27.10 -24.46

yoNhoN -26.55 -30.26 -27.30 -30.66 -25.55 -27.30 -25.75 -26.44 -24.60

gohoN -26.20 -28.14 -27.76 -26.95 -26.46 -27.76 -24.28 -25.61 -25.03

iqkai -27.02 -28.31 -28.74 -26.57 -27.82 -28.73 -27.78 -28.19 -27.88

nikai -27.81 -33.35 -27.67 -27.18 -27.84 -26.62 -26.10 -25.46 -24.22

saNkai -25.03 -31.36 -29.84 -28.56 -32.29 -28.19 -28.21 -26.79 -25.96

yoNkai -25.34 -28.79 -31.68 -28.48 -26.40 -29.13 -26.45 -26.94 -25.41

gokai -26.16 -27.57 -30.64 -26.26 -26.63 -27.89 -24.91 -25.50 -25.98

iqpuN -25.77 -28.02 -25.02 -26.22 -26.97 -26.95 -26.13 -26.75 -24.79

nihuN -27.96 -27.96 -26.21 -26.13 -26.45 -26.10 -26.53 -27.08 -25.02

saNpuN -24.22 -27.81 -26.48 -27.21 -26.03 -29.52 -23.85 -30.30 -24.77

yoNhuN -26.43 -27.35 -27.41 -27.49 -26.77 -27.37 -28.40 -27.56 -25.15

gohuN -26.88 -24.54 -25.46 -26.60 -24.79 -27.60 -27.91 -26.34 -23.86

 

  
図 3.4: HMM scores for M-set

(3)

As for /p/ in [saNpuN], non-native learners E and G did /b/ as we expected that the mispronunciation [saNboN] [saNpoN] or [saNpuN] [saNbuN] would occur.

The use of the absolute threshold function on phoneme score was of little use, as the result showed. However, we found that the automatic scoring based on HMM likelihood score correlated well with human perception.