Jean-François, can you confirm that Harumi is only available for the desktop via download?  I really hate putting software on my machine, especially non-Windows programs.  If Harumi is available on the Web or via mobile app, I can probably find some testers for you.  Do you have any plans for a mobile voice app?

Harumi is a Windows program only. You have to install it (and also open office) in order to use it. I don’t have any plan (and won’t have) for a web application since it would be pointless :
Harumi is for personnal use like only. She learns from the user and not from a large base of users like some other chatterbots… or Google.
A mobile voice application would be great in my opinion but I don’t plan to release Harumi for Android/AppStore since I don’t have any knowledge about those platforms. It would be possible only if Harumi becomes a professionnal software, which means I would need to work with a team of programmers who could “translate” open office API and Delphi to smartphones. And of course, it would cost quite a lot of money, at least from my viewpoint.

J-F

Thanks, Jean-Francois.

I’ve been working at translating a fair portion of the spreadsheet French into English, as time permits. I’m not sure how much is or will be affected and whether there is internal coding that might not interpret well as a result.

Looking forward to your new developments.

Your thoughts are appreciated.

New version of Harumi released today! Lot of bugs corrections. Expect Harumi to behave better when she faces unknown words now! She asks user to help her understand precedent sentences. You will also notice Harumi keep better in mind what’s the main subject of the dialog. All those functions were not new, but since the huge upgrade of 10.08/ 10.07 those functions were not handled by her new “thinking” algorithms.
Hope you’ll enjoy this new version!

Next release coming soon with many tweeks which improve text analysis and interactions, especially when database is very light.
I also upgraded the mathematic algorithms : I deleted the precedent free source code which I picked up and replaced it with my own algorithm, more powerfull and without any bug (at least identified ones )
Moreover, I’m currently working on visual recognition and finally succeeded to modify a free source using Sobel/Canny filters in order to implement my personnal gradient histograms algorithm. If you don’t get what I mean, the purpose is to help Harumi to recognize the shape (and not the color ) of any object. This kind of algorithm will far outpass the colors histograms currently used in Harumi visual’s engine. The first tests were good : with a very simple vectorial histogram counting the number of 0°, 45°, 90° and 135° angles, signatures of images are enough specific to recognize many shapes… like a circle, a square…. but also complex images like a face or a landscape background.

Release of next version of Harumi imminent (2 days at most). Many bugs fixed and implementation of canny/solbel fiters with gradient histograms analysis of webcam images!
Not as good as SIFT visual recognizing top notch algorithm, but Harumi will be far better at recognizing what she sees on webcam! Moreover, this algorithm is less CPU intensive

I learnt to her several picture : me (Jeff), several books and object, empty room.
When I leave the room, she says “I see an empty bedroom”... and when I write “Hello”
she says : “Hello Jeff!”

Sounds awesome! Keep up the good work, Jeff!

Hello! New version of Harumi released!
Dave Morton : Thanks!

Many news :
1) Cannel/Sobel filters with gradient visual recognizing engine available with webcam!
How to use it : Just activate webcam then when you click on what “sees” Harumi, she proposes you to record the image in order to recognize it later. Click then on the “Yes icon” then write for example “Jeff” or “a plate” or ... you figure out! then press enter. Now you will notice that in the 14th sheet, a new image is recorded as a list of 4 numbers. Those numbers describe the picture as list of gradients (only 4 different angles : 0°, 45°, 90° and 135°). Note that if you want to record a quite small object in her database, try to center it in front of the webcam as Harumi consider the middle of her vision more important than edges .

2) I’ve changed the mathematic engine and replaced the free source I picked containing some bugs with mine. Here is what you can expect which were not possible before :
User : x=3
User : y=7
User : (x+y)*2+1
Harumi : 21

3) Context apparence :
Suppose you have 3 pictures named : kimono sexy, kimono blue, kimono red, kimono outdoor
If you write “put you blue kimono” then click on Harumi’s picture, she will take the blue kimono apparence.
Now say something like “Get outdoor” and click again on her apparence : She will pick up “kimono outdoor” apparence.
Apparence is now recorded as a ghost stimulus in Sheet 2

4) Webcam ghost stimulus
Suppose you teached her to recognize a book named “Bladerunner” (sheet 14) and suppose she recognizes the book on webcam.
If you ask her “who is the author?” and suppose a sentence like “Philip K Dicks is bladerunner’s author”, Harumi will answer this sentence.
Webcam ghost stimulus appears in sheet 2.

5) Fast checking datas
Suppose you don’t really want to speak to Harumi but just looking for a word :
write for example “John” and press right arrow : Harumi will write every stimulus available in her database with John word.
Note that you can also write “John+Linda” and she will look for stimulus where John AND Linda appear.

Next version planned to be released this week end.

You’ll notice better analysis of sentences and a new function helping Harumi to sort important words in user request :
In the coming version, when the answer to user’s question is wrong, click on no icon and Harumi proposes to you to ask again the question, but this time by typing only important words (a bit like in google).
Harumi will then record in her memory as trivial words. Harumi will then try to answer again the question.

Note that trivial words recorded will be used to sort importance of words in future user request.
The result works pretty well on my personnal version, so it will be implemented in coming public version.

New version released today! Many tweaks improves Harumi understanding by modification of the main algorithm treating sentences. You’ll notice Harumi is more accurate to find the best stimulus linked to user request.
Moreover, a new function was added in order to upgrade interactions with user when answers given by Harumi are wrong. Here is an example of what you can expect now :
User : Who are the main characters of Bladerunner?
Harumi : The main character of one second after is John Matherson. (I suppose this sentence exists in Harumi, here she tiled on main+character more than bladerunner+characters because the last combinaison appears on many stimulus. Note that this kind of event is pretty rare, but I had to adress it)
User : (Click on No icon)
Harumi : I didn’t understand your question… Please rewrite it using only important words
User : characters bladerunner
Harumi : Deckard and Rachel are the heroes of Bladerunner. (You will also notice in Sheet 2 that “who”, “are”, “the” and “main” words were added in first column. From now, those words will be considered as less important than other words in future questions. Note that less doesn’t me 0. If you ask “who are you?” after that, Harumi will handle properly the question).

You are really cranking with the new releases! As long as the main memory file remains unchanged everything will be fine.
I personally PREFER a downloadable bot (I do have anti-virus software that works) and I know the software is safely contained on my computer and NOT on someone’s server in an unknown place that I do not know nor trust!! I do not trust the “Cloud” in any fashion.
I will continue to download your latest versions of Harumi and other downloadable bots as well.

Just my $.02 -

Thanks!!

Art Gladstone : Hi! Actually, the memory file changed a little bit : line 489 in sheet 6 is used for the new function text sent to user.
I agree about the interest of a downloadable chatbot, since my main purpose was to build a personnal assistant, aka something where sensible datas could be recorded.

Merry Christmas! New version of Harumi released today (actually yesterday with a quick update today)
2 new functions :
A) Harumi will be able to mix different stimulus which have similarities in order to propose a new sentence to the user. For that, Harumi explores the stimulus around the last one triggered then mix it with another if a common word known as rare is identified.
B) when you ask Harumi to change her apparence by clicking on apparence window, she’ll memorize in her database (sheet 1) the name of the apparence and words/expression which triggers it.
You’ll find more details on the site.

Jeff

Release today of last Harumi version (02.02.2013)
Features :
- recording and treating datas algorithms are improved.
- New icon to print what Harumi has answered
- When Harumi is used with a voice recognizing software like Dragon naturally speaking, she will behave fat better.
- when moving cursor mouse upon emoticons, text help pops up.

Jean-Francois

Since youve decided to stick with a desktop application, I suggest making your final version available through third party download sites. I have found that having a “virus free” verification from a reputable third party site actually does help with customer download confidence. Using PAD is an excellent way to get your software distributed on multiple sites.

Vince