developer.* Blogs - My Experience with Speech Recognition: From Yesteryear to Tomorrow (Part 1) - Comments

Speech Recognition: Do Your Homework

Daniel Read — Mon, 27 Mar 2006 04:19:25 -0800

Editor's note: I am closing further comment on this thread since it has unfortunately only become a home for "help me with my homework" messages. (Gunish, I think you could make a nice living helping students with their speech recognition homework assignments.)

All the best,
Dan

Speech recognition with ANN + GA

mapcon — Mon, 27 Mar 2006 03:16:55 -0800

Now i'm doing project for final bachelor course of Comp Sci - it is "Speech recognition based on Artificial Neural network and Genetic Algorithm". I use VC++ for coding, and some main modules are completed, it's enough to demo. But now i want to build a new module that can be draw spectrum frequency of origion .wav file and after applied FFT. Who has source code of above module, plz help me. plz to me by email letamn@gmail.com . Thanks a lot. Mapcon

Learn about Speech Recognition in VB Dot NET

Dawood Khan — Wed, 01 Feb 2006 20:51:38 -0800

Plz sir send me detailed info and some idea how to program for speech recognition in VB Dot NET

speech recognition using VB.net

sarathi — Tue, 17 Jan 2006 17:09:51 -0800

how to write coding for speech recognition using VB.net

i need urgently

irulapparaj — Tue, 10 Jan 2006 01:25:45 -0800

hello sir how to recognize speech from engine in vb.net

Thanks! But how ?

Gunish — Tue, 19 Jul 2005 06:28:46 -0700

hey Garima,

Always Wonder how you never fail to amaze me,There are so many 'Hows ' that i want answered :

How are you?
( Hell i wanna know that )

How did you find me out?
(i though we were never gonna communicate again all of our lives!)

How did u know i got a new job?
( yeah i got a great job, as a Sr. Lead Software Specialist (.Net)
at www.gatesix.com the company is not that big but its a great startup. I gotta Build a Entire .NET Development Infrastructure and Team by my self, phew!)

send me a mail at gunish.chawla@gatesix.com

dont get lost now!

-Gunish

When is the Next Part gonne be out!

Garima — Mon, 18 Jul 2005 06:41:33 -0700

Hi Gunish,
I found this article very intreaguing, i hope that the next part comes out soon, altough the english was not very flamboyish but still it was worth reading,
anyways, best of luck with your new job, hope to read more soon!
Garima

Our (same) boat definately has a HOLE !

Gunish — Sat, 02 Jul 2005 11:07:59 -0700

The reason i chose microsoft technologies over other such as Java is simple, 'to avoid reinventing the wheel'. Microsoft has always provided trmendous support over their own stuff than anybody else, really. and yes dealing with API's is really an complicated issue which certainly requires more than a 'schozmo' developer like me! lol.

Anyways and you are perfectly right about pointing toward Neural Nets as the area under review by maximum developers, weather be it Handwriting recognition or any other human interface parser. The Issue that i have come across is that Neural Nets are STILL not used by Microsoft in doing Speech Recognition in the SASDK's Telephone English Speech recognition Server, nor the ON-NOTE PDA , tablet pc - handwriting recognition algorithm, my contribution being i am close to perfecting simulation over a neural net model of speech recognition patters, i have constantly employed neural nets in various idea over many years from networking to vision parsing but everything is miserable not even near the mark of acceptable range.
Instead of giving up i have tried to use them in many places and out of all the applications that can DIGEST a neural net, speech recognition has been the most promising.
My next part of the article is supposed to talk about how the SASDk works and how this can be considerably improved by employing neural nets to not SYNTHESISES speech but to recognize one ... i am typically using a 3 layer network based on the KOHNEN , GROSSBERG Model of neural nets, its far from acceptable standards but still i do believe that this is how the future models will definitely be built!

About the language recognition... i would recommend you towards a very insignificant product out on the net but with a very effective algorithm. These are called yahoo Crackers, typically work to steal yahoo ID's passwords by many techniques such as brute forcing etc. but since the introduction of the "Enter the Letters Printed Above" scheme in almost all the websites , including Developer Dot Star, hackers have started building algorithms that can read distorted English.. i have definitely come across at least two crackers that were able to constantly Brute Force yahoo by successfully overriding the "Enter the Keywords Printed Above" issue.
the introduction or curvature such as a 3d- ball beneath the x-y alphabet plane introduces a significant amount of distortion in the alphabets which normal recognizers fail to process. This defiantly is a challenging problem, and results of recognition algorithms only get there half way across... the recognition ratio being around just 53%,
i am sure this is what you mean in the Chinese language recognition.
i have plans to work on this issue when i get to the part of implementing Vision in desktops... moreover .... i have once seen a program at The Discover Channel, which showed an Robot At the MIT AI Lab which would 'LEARN TO RECOGNIZE and DIFFERENCIATE' between different shapes such as a Ball or a Cube, this was one thing that was not algorithmically implemented but was done on the base of a Self Learning Expert System clubbed with a neural network... the robot eventually tries to GRAB the ball and if the ball moves away, and it is not able to reach it, it extends its arm.. this also caused the robot to realize the length of Itâ€™s OWN arm, a very interesting situation from AI's point of view other that the classical Sheep Dog Simulation or the PICK up a glass of water and Put it BACK, situation..
at this point i would certainly like to say that there is no PERFECT Algorithmic way to address this problem, thus we should be working more towards 'Writing Programs that can write programs to write Programs' Methodology...
Indeed this is the only way we can Fill up the hole in our BOAT by coming up with a solution that Microsoft is NOT Interests in developing or investing.
-
Gunish

Thanks for an interesting article

Edward G Nilges — Sat, 02 Jul 2005 06:41:08 -0700

I have been pondering a related problem, Gunish: accurately recognizing Chinese handwritten characters.

In the past, I thought the answer was to raise the "level" of the geometry, where geometry consists of "metric" geometry (study of properties such as length which are not invariant under simple changes), "projective" geometry (study of invariants under projection and perspective) and "topology" (study of invariants under continuous, nonbreaking changes in shape).

I thought one might have a shot at recognizing handwritten letters in various languages by considering only topological features. The trouble is that (for example) the English letter A is topologically equivalent to O, therefore you need a separate geometry for "things that stick out like limbs" and "things that come to a point with respect to the rest of the letter".

At that point it seemed that the recognizer would have to be its own mathematician and dynamically evolve mathematical theories for recognizing variations of a letter.

It would in fact have the same sort of problems I faced in learning to write Chinese, where I would make embarassing blunders in calligraphy on the new Shenzen subway, pointed out by six year olds. As it happens, for example, the second bar of the symbol for the number three has to be shorter than the first bar.

Yet, when you examine signage in Hong Kong, the letters are systematically distorted!

Where we've gotten to with respect to both voice and handwriting recognition is the idea of "training" the machine, but consider some kid dumped by his parents at Lo Wu in 1955, who grows up in Hong Kong, and learns to read signage because it's that, or starve.

He's forming scientific theories and testing them.

A teacher of mine at Princeton, Gil Harman, did work in Lisp on models that would form theories and test them but as far as I know this work is commercialized only the form of neural nets.

Anyway, certainly sounds like you are on track to something. My only warning, apart from the above considerations, is that relying on Microsoft APIs to be stable over multiple releases of Windows is a bad idea. Microsoft's policy in the past has been to adopt favored companies, and give them information on changes to APIs, and new APIs, only on condition that they become official Microsoft certified sites.

This is why there are no API calls at all anywhere in the software for Build Your Own .Net Language and Compiler. Instead, the utilities.DLL and windowsUtilities.DLL do the best they can to provide needed functionality exclusively in terms of the documented, and presumably stable, behavior of Visual Basic .Net.

This has its own dangers. I use "legacy" character input and output, for example, in file2String and string2File to translate files to and from strings. There's a chance that in some future release, the functionality exposed by Microsoft.VisualBasic may in some cases be downgraded out of existence if it is inconvenient to support.

However, wrapping the function in a clear, transparent, Saran wrap like "file2String" makes it obvious that this is the goal, and, the code can be replaced.

Since Build Your Own, I have started always qualifying ANY reference to functionality that is in Microsoft.VisualBasic with "Microsoft.VisualBasic.x" so as to be able to find possible exposures. Soon, it shall be time to simply remove the reference and the imports even from VB.Net projects because in many cases the functionality is available in more reliable (and, more internationalizable) form elsewhere.

Or, to junk VB. C# is .Net "equivalent" but VB encourages legacy and USA centric ways of thinking.

But, I have decided, not before a VB version of Spinoza is complete, because it is quite late in my own personal game, and developing a compiler for a new language is my highest priority.

You and I are in the same general boat. You are relying on Microsoft software, and I am relying on VB.Net having a future in the global marketplace.

However, my experience has long been that you can drive yourself batshit by backing up continually to do things "perfectly".

hi, u must be the same guy w

Swetha Medhani — Sat, 02 Jul 2005 00:44:37 -0700

hi,
u must be the same guy who i met at OIST Software expo... because there also yopu had brought an AI Project that was working in human interaction through speech,well i asked u there itself how did u make it and u said 'its a long and booring story' i guesses you didnt want to talk then. i had decided that time that one day i would find you and make you teach me how to work around with AI and speech in particular,so i guess i HAVE found you and now i want you to help me out, post your contact number or something so that i can call you and get in touch with you!
-
Love ur work
Swetha

lot of experience with speech

Pooja — Sat, 02 Jul 2005 00:39:49 -0700

lot of experience with speech recognition i see!
when did u start working on it anyways
... i think i met you at a software contest hel at NRI Institute ...
and i am sure that you dont remember me ... but i was preety impressed by your imagination and futuristic vision.
i hope that you post the next part soon enough so that i can get the complete idea of it !
-Pooja

hey this is ankush,, nice o

Guest — Sat, 02 Jul 2005 00:37:10 -0700

hey this is ankush,,
nice one ... was very informative .... i hope you will be writing the next part soon though!

My Experience with Speech Recognition: From Yesteryear to Tomorrow (Part 1)

Gunish Rai Chawla — Wed, 22 Jun 2005 11:06:45 -0700

I am in no way suggesting that voice recognition systems have become perfect, but what I am implying is that we have started to tread the right path...