Comments for B.Log.Likelihood

Comment on A more human voice by Mahdi

Mahdi — Thu, 23 Feb 2012 14:57:26 +0000

Thanks for the comment Ehsan. Jordan Frank also suggested the use of sentiment analysis for this project. I’ll be looking into that as well. For now, I’ll probably focus on HMM type modeling of the tone and speed over long texts. I haven’t gone too far with the project though. Very busy with PhD, TAships and ICML.

Comment on A more human voice by Ehsan

Ehsan — Thu, 23 Feb 2012 02:09:14 +0000

Hi Mahdi,
just out of curiosity! have you ever thought about using sentiment analysis over sentences to add more general informations like the tone, feeling of a sentence? In text processing you can’t go too far by just using POS tags. Even if you want to use grammatical structures like parse trees, still you need something else to use on top of it.

Comment on A more human voice by Well-Documented Hobby » B.Log.Likelihood

Well-Documented Hobby » B.Log.Likelihood — Sat, 05 Nov 2011 12:36:58 +0000

[...] the note previously posted here, I have started a website to discuss and exchange ideas on the realization of a more human-like [...]

Comment on A sad ending to a successful social experiment by Saqib Ali

Saqib Ali — Fri, 21 Oct 2011 15:07:46 +0000

Mahdi,

Interesting observation. In retrospect, if I think of my Google Reader usage pattern, I have been using it exactly like you described. I don’t use Facebook, Google+ etc. In fact I don’t even like those platforms because of the unnecessary information overload they create. Whereas, I have been actively using Google Reader since the day it was released. And the reason, as you described, is that I don’t really care about the personal lives of the people I follow on Google Reader. I just want to know what they are reading. It is like going to someone’s place and browsing their library and not being concerned with their personal lives. I like that pattern.

I too hope that Google Reader lives on in its current form.

Saqib

Comment on A more human voice by Pablo

Pablo — Tue, 04 Oct 2011 12:30:13 +0000

I think that even if you surpass all the technical difficulties with this learning problem, there is still a limit as to how suitable and/or natural the speech will be. The intonation given to a sentence is largely subjective and can change the meaning of the sentence completely. Irony can be used, through intonation, to turn praise into ridicule, and the intention is a part of the speaker. Regardless of whether it is the speaker’s own beliefs, speech conveys information on more levels than, say, uniform-font print would.

It would be interesting to have a system that can receive input from a user as to the desired intent of the passage, producing different “readings” of a passage based on this intent.

Good luck with this project!

Comment on Note to Self – Honest Research by Csaba

Csaba — Wed, 28 Sep 2011 00:43:16 +0000

very well sad!

Comment on A more human voice by Mahdi

Mahdi — Wed, 28 Sep 2011 00:27:16 +0000

Thanks a lot for the comments Csaba. I certainly am not going to start from scratch on this and will most probably use a commercial engine here. I will try to tune it using pause and speed “control tags”. I will not be generating any new voice, per say, and will only try to make the existing ones sound more natural. That should wash out a whole lot of details, so that the focus can be only on a few parameters to tune rather than the larger scale problem of speech synthesis.

Comment on A more human voice by Csaba

Csaba — Wed, 28 Sep 2011 00:07:11 +0000

oh, this was like 15 years ago.

Comment on A more human voice by Csaba

Csaba — Wed, 28 Sep 2011 00:06:44 +0000

I was working in speech synthesis for 5 years. In the high time I had 5 researchers under my hand and a bunch of engineers. We got some good successes. At the time Rhetoric, AT&T, Learnout & Hauspie were the big players. Then I got back to do other things. One thing I have learned is that simple things take a lot of time. And high quality data does matter. And that text-to-speech is in fact AI complete. We used all kind of machine learning to predict all kind of aspect of speech, mostly surface techniques (prosody, rythm are extremely important and if you get them wrong, they can easily give away that a computer speaks). Low level synthesis at the time was already pretty good, the only problem being that it was too expensive to create new “voices”. If you want to do this at a hobby level, start with a good base and try to improve some part of an existing synthetizer. The whole codebase of our project was probably close to 1M lines.. Good luck!

Comment on A more human voice by Jordan

Jordan — Tue, 27 Sep 2011 14:42:50 +0000

Great post. I’m starting to work on some NLP stuff for my research, so if you ever want to chat about this, or want help with your new hobby project, let me know.