On the slim chance that someone here wants to re-implement the unmodified Kneser-Ney algorithm[0], the presentation of it by the book does not account for unknown tokens in the query not in the vocabulary. I extended the recurrence to its natural closure including unknown tokens here [https://github.com/jhanschoo/HMMTagger/blob/master/readme.pd...]. A straightforward task, but it might take you an hour or two (probably more) otherwise to obtain it and prove its correctness, seeing as I couldn't find an extension in a Google search nor is it described in the original paper as well. I believe that it would likewise be straightforward to extend this to modified Kneser-Ney as well.
[0]: The modification of using multiple discount values due to Chen & Goodman is regarded as a more well-behaved smoothing, and more popular today.
The webpage https://web.stanford.edu/~jurafsky/slp3/ links directly to the PDF, and gives context and other download options. It's not so easy to go back from the PDF to the webpage.
Mods: I'd change the link, and the title to "...3rd Edition draft".
Everyone else: please stop linking to PDFs when there is an obvious html page to link to instead.
I found speech and language processing to be one of the most interesting courses of my degree. Recently I decided to take a look at speech synthesis again and discovered a book by Paul Taylor on this subject (http://svr-www.eng.cam.ac.uk/~pat40/ and draft PDF at http://svr-www.eng.cam.ac.uk/~pat40/ttsbook_draft_2.pdf). It is more engineering focused than other books in this area.
Note that the book was released 2009 and the draft is from 2007. While most of the content is still very relevant for understanding the basics and challenges in building TTS systems, recent progress in DNN-based synthesis (including WaveNet, GANs, and end-to-end approaches like Tacotron) is obviously not covered.
Often this stuff is used for surveillance. If you choose to study this area, please be careful how you apply your knowledge. There are plenty of positive ways to use it: contributing content classification systems to sci-hub or libgen, building tools for the disabled, automating multilingual visual design aesthetics with computational linguistics and machine learning...
Whilst I understand the concern but just Internet is now for control and tracking (chi-na is more obvious but post someone we all know better), we cannot stop the wheel.
We need an equivalent of the hippocratic oath for software engineers: I swear by Hephaestus, the ugly god of crafts, that I shall not use my skills for controlling and tracking ...
Early in my career (it began after a discussion in a tutorial during my undergrad) I made a choice to never to use my engineering skills to build weapons. I don't believe its a morally justifiable choice in this century. It's not like I have been forced to starve as a result but it did reduce a few opportunities.
Great, these who want to build these things will simply hire people who don't subscribe to this oath ... And now the organization will be populated by a whole tribe of people who have no qualms about doing this -- does that sound like something that will make things better?
Aside: the book looked interesting. Skimmed it, beautiful typesetting, not gratuitously mathematical, looked readable -- I may be wrong, but it's just a exercise of adding awareness of it's existence to me.
> Great, these who want to build these things will simply hire people who don't subscribe to this oath ...
How many doctors are willing to give up their title to do unethical stuff?
An ethical code will change people's perception, and that's a big thing. It will morph the perception of Google and Facebook from cool tech companies into dark places where no decent people want to work; unless of course these companies are willing to change. And there will be a huge effect on politics, other businesses, and media.
Um, how are they supposed to make money if they don't track you to a certain extent? Do they have flaws? Yes -- it's not funny what kind of things you can target for based AdWords, their internal staff probably aren't creative enough to think up ways in which their tools can be used to inflict misery. Software engineers aren't soldiers, doctors, or civil servants what kind of ethical code do you expect? How will people be held accountable to it? What will happen to these who make transgressions?
> Great, these who want to build these things will simply hire people who don't subscribe to this oath ..
Morally, this argument quite obviously fails because it could be used to justify anything.
Practically, we see many professions where people of character have successfully resisted temptation: every government employee who is not corrupt comes to mind.
And even if they'll just hire the next guy/gal: If you say no to an offer, you were, by definition, their top choice. That next one down the ladder will therefore be slightly worse, making such businesses less profitable.
Top choice based on a search with finite resources, maybe there is a better top choice if they spend a bit more time -- one that is also more agreeable. I'll return with a question -- Is resisting temptation the path of least resistance? What are the odds that the personality trait you describe also overlaps with the person being in a position of enough authority to actually matter?
NLTK tutorials are interesting for the absolute basics in Python, e.g. http://www.nltk.org/book/ch01.html. The field has moved into advanced Bayesian and neural methods, but NLTK still has a use in pre-processing and thinking about how to work with text. The first thing to recognize, I think, is to realize how hard many of the language tasks are that people do trivially.
I need a voice activity detection module (VAD) for my wearable computer. Should I roll my own, or use someone else's (open-source). My immediate need is speaker-dependent (just me), but it would be nice if I could offer up a speaker-independent version eventually.
Thanks! This is exactly what I needed. I was able to get it up and running on Raspberry Pi with just "pip3 install webrtcvad", and the quality is at least good enough to get me started.
[0]: The modification of using multiple discount values due to Chen & Goodman is regarded as a more well-behaved smoothing, and more popular today.