This is the result of a well known devoicing rule in Japanese. Devoicing means that there is no vibration of the vocal folds. For example, the difference between [s] and [z] is only that [z] is voiced. The IPA diacritic for devoiced phones is a circle at the bottom of the glyph e.g. [z̥]=[s]. Although there is still much dialectal, idiolectal (the way a particular individual uses words), and lexical (the way a particular word is used by individuals) variation in devoicing vowels, and in fact, what is called Japanese vowel devoicing does not in all cases result in merely a devoiced vowel. Taken from The Sounds of Japanese (Vance, 2008):
The so-called devoiced vowel is actually missing entirely in many cases, although traces of it remain as coarticulations in the immediately preceding consonant. Some researchers refer to the affected vowel as reduced, and this term is more accurate as it covers a wider range of possibilities.
A coarticulation in a preceding consonant means that, for a consonant vowel sequence the consonant has been pronounced (articulated) slightly differently in anticipation of the upcoming vowel, so different vowels might result in slightly different articulations of that consonant. Coarticulation is one acoustic feature that complicates speech recognition/synthesis.
The traditional and most simple description of vowel devoicing, taken from An Introduction to Japanese Linguistics (Tsujimura, 2006), gives two conditions for devoicing which Vance calls a good first approximation:
- The high vowels /i/ and /u/ are voiceless when they are at the end of the word and are preceded by a voiceless consonant
- The high vowels /i/ and /u/ are voiceless between voiceless consonants.
The う in です# and ます# (# represents end of utterance) satisfy the first condition and so undergo devoicing:
- desɯ# → desɯ̥# since /s/ is a voiceless consonant
- masɯ# → masɯ̥# since /s/ is a voiceless consonant
If, for example, we have ですか, the う is still devoiced because /ɯ/ is found between /s/ and /k/ (/desɯka/) which are both voiceless consonants and so the 2nd condition is applicable. Devoicing a vowel can be tricky if you deliberately try it in isolation, but to do so just don't let your vocal folds vibrate.
However Vance offers the following relevant observation:
Devoicing between a voiceless consonant and a pause is much less consistent than devoicing between two voiceless consonants.
There is one immediate complication (among several) to devoicing high vowels; what happens when a vowel is both devoiced and pitch-accented, as in the /i/ of 四季 /ɕika/? The contradiction is that you cannot have a high pitch on a phone that isn't voiced (vocal fold vibration, or phonation, is what generates that part of the signal that is perceived as pitch). Usually textbooks (or at least the two that I know of) in a first course of Japanese linguistics will not address this contradiction. This paper Against Marking Accent Locations in Japanese Textbooks (Hasegawa, 1995) argues against this pedagogical simplification:
The fact that native listeners do hear an accent on a devoiced syllable indicates that associating an accent invariably with a high pitch cannot be an accurate description of the language. This paper discusses how Japanese accent is actually realized and argues that marking accent locations in textbooks without a detailed explanation about accent is merely an extra complication that introductory textbooks should avoid.
Devoicing becomes rather complicated when intonation and other exceptional cases are considered, but just looking at the sheer number of papers on the topic it seems to be a well studied and documented feature of Japanese phonology.