I've got an old PDF folder full of papers on Japanese, and I managed to pull up two which might be helpful. (I've been on the search for a full detailed phonetic study of Japanese. Add a comment if you know of some other technical resources!). The first, the open paper Processing missing vowels: Allophonic
processing in Japanese (Ogasawara and Warner, 2009) (from the journal Language and Cognitive Processes) and the second, which I think needs to be purchased, Vowel devoicing and the perception of spoken Japanese words (Cutler & Otake & McQueen,
2008). If you can take a look at those papers, they have spectrogram of the voiced/voiceless sounds and they are quite informative.
Ogasawara has a very informative diagram at the beginning of his paper which displays an oscillogram and spectrogram of two phonetic variants of the non-word /hokita/; one in which the speaker obeys the devoicing (reduction) convention, and the other in which the speaker diligently avoids the devoicing tendency. So, two phonetic forms [hokʲi̥ta] and [hokʲita]. The observation made is that the reduced vowel has no periodic wave or frequency mass resembling a formant, so that the vowel [i̥] appears to be virtually deleted. What is found after the [kʲ] in [hokʲi̥ta] then is "frication noise following the plosive burst of /k/, so that the vowel [i̥] consists of only low amplitude palatalized voiceless noise."
The paper is free, so take a look two spectrograms corresponding to the [-voiced] and [+voiced] forms at the bottom of the figure. The characteristic voice bar which indicates frequency mass at low frequencies (characteristic of phonation) are absent in the devoiced vowel. So, clearly there is some veracity to the term "devoicing". For the [-voiced] vowel, I can't even see any noise in the waveform, so it appears as if [kʲ] is just an extra prolonged stop. Which seems in agreement with Ogasawara's opinion that "phonetically, so-called devoiced vowels are often deleted (Vance, 1987, in press; Yuen, 2000), or a short, low-amplitude vowel may remain (Yuen, 2000)."
It seems that some some authors (Ogasawara, Vance) call it reduced instead of devoiced to skirt the debate between devoicing vs. deletion. However, that is not the end of the phonetic story for the unvoiced vowel, as it has a coarticulatory impact:
Japanese vowels, both unreduced and reduced, cause coarticulation in the
preceding consonant, which allows identification of the /i/ or /u/ even if the
vowel itself is deleted (Ostreicher & Sharf, 1976).
This suggests that the best way to recognize a devoiced vowel is to look for consonants bearing the acoustic cue of this coarticulation as opposed to measuring and comparing for "low amplitude palatalized voiceless noise". Importantly, Ogasawara points out the simple reason why on phonological grounds a vowel must be there in the underlying form:
The reduced vowels are considered to be present at least in the underlying form, because they cause coarticulation, and because, if they were not present at all, this would leave consonant clusters (e.g., /kt/ in [k(i)ta] ‘North’) that are otherwise
phonotactically impossible in Japanese.
And other interesting ideas/observations:
Cutler et al. (in press) find that listeners do not restore reduced
vowels at an early, automatic stage of processing. Furthermore, they find
that the number of words in the lexicon containing a given string with a
reduced vowel affects how likely listeners are to assume a reduced vowel is
present. They conclude that Japanese listeners’ restoration of reduced vowels
happens during lexical, rather than prelexical, processing.
Reduced vowels are acoustically weak, which might make them
harder to process. However, phonotactic knowledge (which indicates that a
vowel must be present because of the consonant cluster) should facilitate the
recognition even of reduced vowels. Moreover, language-specific knowledge
of the allophonic alternation should facilitate recognition of reduced and
unreduced vowels in their appropriate environments (e.g., [(i)] in [k(i)ta]
‘North’ and [i] in [itRigo] ‘strawberry’).
Although devoicing is not obligatory, analyses of
the Corpus of Spontaneous Japanese ͑Maekawa, 2003 show
that it is highly probable ͑over 98% in some environments;
Kondo, 2005; Maekawa and Kikuchi, 2005.
And on the perceptibility of the devoiced vowel, Cutler says:
The effect of this devoicing is the creation of sequences
of consonants not separated by the periodic articulation
normally associated with vowels. In contrast, insertion of
a vowel into a consonant cluster ͑e.g., fillum for film͒ makes
recognition easier, in part because the consonants in the clus-
ter indeed become easier to identify if separated.
So, this might not have been all the information you were looking for, but beyond this you start to get into statistical analysis of signals and techniques from experimental design, which I don't fully understand.
