truthexposer
Speech models usually use triphones, which turns out to be a huge amount of audio. This is particularly impressive because of how little data they need.
Google used their own datasets, which are most likely massive.
Even then, I don't believe the issue is with stress. I believe that the voices sound robotic because they are using, and also admitting because it makes their results impressive in some sense, very few samples, "less than a minute" they claim. Triphones are usually what speech systems are trained on. The amount of triphones (3-phoneme-grams) to cover a language's phonemic inventory is huge (50 phonemes = 50! triphones, which could mean a few hours of audio, although many will not occur within the language given the phonotactics of the language).
tell the eskimo seeing X different kinds of snow to any linguist and they will roll their eyes, debating whether or not to explain why the example is probably not only incorrect, but fabricated for the umpteenth time.
And with respect to the phonetic shifts, from a linguistic perspective, most changes of the phonetic radical involve one change of the initial sound, i.e. bilabial to interdental, dental to alveolar, voiced to unvoiced. With respect to the phonetic inventory of the language, these shifts are only one feature shifts within an old SPE framework. These small featural differences, whether conscious or not, (usually unconscious because we acquire languages during our infancy) are picked up and used by the speaker of the language to categorize words.
And while of course every language descends from another, and ancient Chinese isn't understandable to a modern Chinese speaker, there is a long tradition within the Chinese language, and change within the language hasn't been shaped by things as drastic as displacement of the linguistic community, or forced political rule under a foreign power, etc.
The Sapir-Whorf hypothesis, which underlies Arrival, has been debunked for years in academia, and the movie, if detached from that hypothesis, is more of a thought experiment on the effect of the linearity of time on human thought than a thought experiment on anything related to language itself.
First, literacy isn't completely related to the writing system. Look at Spanish speaking countries, where the alphabet is more phonetic than the English alphabet.
Second, Chinese characters that are more complex, i.e. consists of more than one radical, are usually composed with a semantic component, giving indication to the character's meaning, and a phonetic component, which gives an indication to the sound of the character. Although this isn't a rule, it helps a lot, and it's not like English doesn't have crazy non-phonetic spellings as well (how tf is "through" supposed to be pronounced for a English learner?)
Last, the Chinese language consists of MANY homophones. This isn't necessarily a bad thing, and not one out of design, but something that is the result of being one of the oldest language families in the world. It allows for the concise expression of many things using only single syllables. You might say, but what about the crazy amount of ambiguity if the language has a lot of homophones? Well, ambiguity is a huge problem in all languages and our brains seem to manage. Now, even though homophones aren't a big problem in spoken language, because of intonation and prosody giving a clue to how to analyze sentences, written language is a different story, and it would be very hard to make an easier system to handle it. For all you engineers, the fact that Chinese has characters is essentially a performance trade off. More information density for more ambiguity.
Here's some trends noted from the 2016 NCDT, which they do admit have only appeared within the past few years: "Behaviors that are slowly growing include lighter coffee consumption (slightly fewer drinkers and slightly lighter cups per drinker per day), drinking espresso-based coffee and drinking coffee out-of-home. Behaviors that are slowly declining include drinking traditional coffee and drinking coffee at home. Note that most of these shifts are occurring over the last few years."
Some other figures based on satisfaction of brewing methods found later in the report include:
86% very satisfied/satisfied with their drip coffee maker in 2016
94% with espresso
90% with single-cup
89% with instant
88% with coffee vending
It seems to me like there is a growing trend of more conspicuous consumption, as well as already high satisfaction across the board with people's status quo brewing methods.
It also seems to me like you're trying to make an instant coffee that carries with it the sort of branding that Philz, Blue Bottle, Stumptown, Starbucks, etc. have with physical locations which have the ability to foster certain groups of people, as well as convey aesthetic or ethical sensibilities through branding within the stores. It seems to me that this kind of branding is somewhat reliant on physical locations. I just don't buy into the assumption you make that you can target the same consumer that these places can.
Like, come on, the title of the NYT review on your coffee is: "Instant Coffee You’ll Actually Want to Drink", if that is any indication as to the mismatched conceptions people already have about instant coffee. It seems like you're climbing an uphill battle, considering the fact a cup of any other regular coffee costs about 7 cents, you're trying to sell instant coffee, but 36 times more expensive, based on branding and convincing people
And if someone is buying instant coffee anyways, I doubt they care about the taste, or have preferred methods of milk/cream and sugar that would make any drown out any nuanced/bad flavor of ANY coffee.