Slashdot Mirror
Google's DeepMind Develops New Speech Synthesis AI Algorithm Called WaveNet (qz.com)
Artem Tashkinov writes: Researchers behind Google's DeepMind company have been creating AI algorithms which could hardly be applied in real life aside from pure entertainment purposes -- the Go game being the most recent example. However, their most recent development, a speech synthesis AI algorithm called WaveNet, beats the two existing methods of generating human speech by a long shot -- at least 50% by Google's own estimates. The only problem with this new approach is that it's very computationally expensive. The results are even more impressive considering the fact that WaveNet can easily learn different voices and generate artificial breaths, mouth movements, intonation and other features of human speech. It can also be easily trained to generate any voice using a very small sample database. Quartz has a voice demo of Google's current method in its report, which uses recurrent neural networks, and WaveNet's method, which "uses convolutional neural networks, where previously generated data is considered when producing the next bit of information." The report adds, "Researchers also found that if they fed the algorithm classical music instead of speech, the algorithm would compose its own songs."
You reference 3 different nationalities - how is that a double standard?
Political Correctness is such a hot mess. It's a shame how the good intentions behind it have been strangulated with censorship and agendas.
You are not a frictionless sphere at rest on a perfect plane in a vacuum. Surprise!
Unless you are somewhere on the autism spectrum or being willfully obtuse, figuring out why people do things is not usually very difficult. As a fun exercise, try and figure it out for yourself using clues from history and culture.
The only thing necessary for the triumph of evil is that good men do nothing.
Political Correctness is fascism with manners
while more natural speech synthesis maybe useful to create an illusion of intelligence, speech synthesis and so called "artificial intelligence", are too different things.
even more relevant learning to mimic speech and word use of others, is just another way of using "artificial intelligence".
Not yet, but I'm working on it.
You are welcome on my lawn.
It's three double standards, so I think it should be 3^2 standard. Though I suppose you could make a case that it should be binary, so 11 standards.
You are welcome on my lawn.
And why isn't Slashdot taking comments on Facebook Features 9/11 Conspiracy Theory as 'Trending' ???
Because the official government conspiracy is real! OPEN YOUR EYES AMERICA.
Listening to the two tracks, I have to say I'm impressed.
And the alt-right are fascists who fetishize incivility.
Pick your poison.
It sounds more natural than a crappy sounding text-to-speech system.
In other news, Apple out-did itself by announcing the best iPhone ever.
> Please explain....?????????????/
Context.
Hardest thing for nerds to figure out is that context almost always matters more than the words themselves. If it helps, think of it as a lookup table - the word itself is just a key to look up much more information, and just as importantly -- few people have identical copies of that lookup table.
Dutchman and Englishman were never derogatory terms. Chinaman frequently was. That's why.
Did they confuse recurrent neural networks and convolutional neural networks when discussing the old versus new method of speech synthesis?
Flouting ./ tradition, I actually listened to Deepmind's examples of their voice. They're rather unimpressive compared to the other two voice samples they compare themselves to, and very noisy. I heard much better from IBM Watson four years ago.
Methinks Deepmind published too soon.
We've come along way from Creative Labs' Dr.Sbaitso.
Anyways, to answer your question. It is the same reason why we call the countries Neder-Lands, Eng-Land, Scot-Land, Ice-Land, Fin-Land but not Chink-Land.
I'll never understand why Slashdot likes to link to poorly written and misleading summaries, when the original blog post is so much more readable and informative. I suggest everybody skip the "Quartz" article and instead read the original blog post. Thankfully, for once it was in fact included in the Slashdot summary, even if it was downplayed: https://deepmind.com/blog/wave...
While the individual words are are better... the sentence pacing is not.
This is similar to the "singing computer" pronunciation, many years ago, in which the ACM distributed CD's with the tracks on it.
You don't get the stilted words, but unless it's intentionally paced (for example, a real human would have put a pause before "directed"), it's still going to be recognizably artificial -- but worse than that: difficult for a human expecting the pacing to understand.
Given that age related hearing loss tends to cut out vowels and not consonants, this is unlikely to be a useful implementation for care giving of older people, for example, unless there are also visible facial cues associated with it, if the pacing can not be made distinct.
It will be great when games will be able to use non-pre-recorded speech for dialogs. No need to have characters express just two or three different game states with one recording each.
Ezekiel 23:20
. . . Mycroft is on the line.
The word is that Star Trek: Discovery may attempt to use Majel Barrett's voice for the computer, due to her having recorded a complete phonetic sample before she passed. If this really does outperform the best available TTS engines, then perhaps DeepMind would be a good fit to generate that for the show: since it's supposed to be a computer, it's not the end of the world if it doesn't sound completely human...
[nt]
File under 'M' for 'Manic ranting'
Political Correctness is such a hot mess. It's a shame how the good intentions behind it have been strangulated with censorship and agendas.
2009 called and they want their "hot mess" quip back.
WaveNet examples seem to have been produced using higher bitrate samples. Otherwise the Parametric examples sound better to me. Either way, if the sample rates were different, the comparison isn't completely fair.
You're right, Star Trek mythos should include the Majel voice becoming a standard computer voice sometime between TNG and TOS, and Discovery should include a voice which sounded machine, like in TOS, probably not yet standardized, so a different voice could be used.
When you have 12 rules based speech synthesis system (around 1967), where given a phoneme (or say a symbol p -bilabial voiceless plosive, has a noise spectrum around 4000 Hz) followed by say, i, to produce the spoken syllable pi, then the sound 'i' has three frequencies 220, 1900, and 2700 HZ, the rule says that for the sound p give a duration of 20 milliseconds, the noise be around 4000 Hz, the lowest frequency of the sound 'i' that follows p should start around 170 Hz and within 30 mmiliseconds reach the target frequency 220 Hz so on. Thus the frequency- energy or voltage-amplitude for each pitch period has to be computed and played back at say, 8 bits per second decoded into analogue waves. Therefore, the rules embedded in the algorithm based on AI tells the computational part to use a table lookup part, retrieve the relevant coefficient for a digital filter for generating a wave with embedded frequencies, synthesize the digital wave and decode it into an analogue wave via a sound card.. The setup is complicated. To this if you add the fundamental frequency of 80-160 Hz for males, 140-280 for females and up to 400 Hz for a child, the rules become complicated. Tone, gender, age and other factors will end up in producing more complexity. Thus, dialect, region, mother tongue, sex, age, abnormal vocal tract compensation, nasalization and other linguistic features need AI at the source of generation. Thus, IBM had voice type dictation which worked only for about 98% time for standard American English. Blacks, Jews, northerners, Indians and Chinese had less than 80% accuracy from the system. If Google had achieved 99% accuracy based on Baye’s theorem and predictive filtering etc., for the general population, then it is a big break through. Time will tell the results.