Slashdot Mirror
Open Source Speech Recognition - With Source
Paul Lamere writes " This story
on ZD-Net and this recent story
on Slashdot
describes the recent open sourcing of IBM's voice
recognition software. This release, unfortunately, doesn't include
any source for the actual speech recognition engine. Olaf Schmidt, a
developer on the KDE Accessibility Project ,
is quoted as saying 'There is no speech-recognition system available
for Linux, which is a big gap.' In an attempt to close this gap, we
have just released Sphinx-4,
a state-of-the-art, speaker-independent, continuous
speech recognition system written entirely in the Java programming
language. It was created by researchers and engineers from Sun, CMU,
MERL, HP, MIT and UCSC. Despite (or because of) being written in the
Java programming language, Sphinx-4 performs as well as similar
systems written in C. Here are the release notes and
some performance data."
Despite (or because of) being written in the Java programming language, Sphinx-4 performs as well as similar systems written in C.
Im sick of these comments. Anyone that needs to know about the performance of Java knows its very fast. Why bother commenting about it anymore?
Its like saying "... and because it was written in C, its very fast...", as if we didn't know already.
I've used a few packages for speech recognition but none really got me too excited. Well, Dragon Naturally Speaking did have me read a few chapters of Dave Berry to it. I bet it didn't work because of all the laughing, I was in tears.
I must say though that speech recognition is something that the whole computer community needs to work on. Now, we can finally do that. All the "open source community" needs is source that works a little. In a year or so, I bet this works better then most options available today.
Now, I know that isn't the rule but this is the type of thing that computer/math engineers could sit down to and contribute where others can't. It seems to be the rule that the really smart ones tend to work with open source software...
Really the cool thing is that this could get people involved who otherwise wouldn't because they don't know where to start.
Get your Unix fortune now!
I disagree. Many of us are not native English speakers. What may be obvious to you phonetically doesn't have to be obvious to the rest of us. :)
;)
This being said, yes, that the explanation of the joke would be modded up is a bit sad, alright.
I could easily live with 10-15% slower, IF Java didn't have the startup overhead. I can run inetd-style fork-exec-terminate servers in C on CPUs that a cellphone would spit on, and handle hundreds of connections a second. Bringing up a JVM on the same processor would take minutes. Bringing up a JIT runtime would be out of the question.
For applications where you can create a JVM and use it as you need it, Java's great. Webservers, sure, no problem. Desktop applications, heck, the GUI overhead's getting to be the same order of magnitude (though that HAS to change, we can't afford to depend on Moore's Law much longer unles someone comes up with a clever way to cut the power consumption of processors faster than the speed increases). Browser plugins? For content, yes, but not for navigation... if it takes 10s to start up a JVM your customer's already hit "back".
It takes at least 3 people to make for a really good joke:
-
One person to tell it
- One person to get it
- One perdon to laugh at for just missing the whole point.
#3 is a little bit less obvious when the joke telling is online.Sometimes boldness is in fashion. Sometimes only the brave will be bold.
You are, of course, perfectly correct in everything you said.
There are a number of HCI aspects where speech recognition is not a good solution.
However, let me enumerate a number of other ones, where it's superior:
Minutes of meetings, or similar. Imagine having a verbatim record of a discussion there by the time you get back to your desk.
Someone who cannot type - e.g. no hands. Rare, granted, but still a viable use.
Someone whose hands are busy. The cannonical example here is a pathologist doing an autopsy, where they dictate everything. Speech recogition saves time in transcription (and money for the audio typist).
I'd love to be able to issue voice commands to a computer, for a few, isolated cases. For example, diagnosing hardware. Bring up a doc, and be able to get the computer to flip pages, without having to remove the probes from the hardware. Re locating them is a pain, and sucks time.
Moreover, I'm certain that there are others, some of which will only be realised when it's common and cheap enough to be widely available.
It's like a mouse. It's one of the worst general purpose input devices for a computer [0], but it's excels at indicating a single element on a display. The mouse and keyboard complement each other, and there are a bunch of other, more specifc input devices, such as the graphics tablet. I have no doubt that if speech recognition was as accurate and reliable as a graphics tablet, it would get a similar amount of use.
[0] Try inputing a block of prose with only a mouse. Even specilist software makes it only suck marginally less.
Heh...you could substitute "uses Linux" for "written in Java", and you'd have the same thing.
Seriously though, Sphinx-4 is really worth looking at. That group at Sun does great work.
With that said, you can probably guess I have a lot to say about Speech Recognition. (Not Voice Recognition, that's different, that would be able to distinguish Ben from Charlie for example.)
A good SR engine is, of course, essential. And I've not read the details on the two recent giveaways, but I suspect that they are only the engine.
The SR engine is just a begining. There is a ton of UI work that needs to be done. Sit and think about spacing around punctuation marks and then think about capitalization around puncuation marks. Yeah, it is all pretty cut and dried and known but the details really need to be sweated to get it right. This is very time consuming.
Next you have to worry about exactly where you are editing. Is that into Microsoft Word (or Open Office), or emacs, or where? It can make a huge difference when you want to go back and correct misrecognitions. You just don't want to send N delete characters and retype it, that results in a lousy user experience. So just exactly where is the input cursor at all times? This is not an impossible problem, but one where the details must be sweated.
Next is command and control. Just how are you going to let the user grab the text of all the menus and all the text in the dialog box buttons. Again, not impossble, but more of those pesky details.
Finally, is your SR engine good enough? Maybe, maybe not. Let just say that 98% accuracy might look good on paper, but that is one in 50 words wrong. Unless your correction mechanism is smooth, an error rate that high greatly slow you down.
Is Open Source SR a good thing? Oh yes sir, yes! But lets not forget the details. One thing the Open Source community has been accused of, perhaps justly, perhaps, unjustly, is not sweating the details.
Speech Recognition has an awful lot of details.
I was thinking about this the other day, and was wondering if this is a huge gap in the Windows user interaction model.
Think about how you input info using windows. You click on a few locations using the mouse, perhaps use some keyboard input, click some more. The output from these inputs is arbitrary: it may result in anything from a 'File/Save' dialog to a custom error dialog box. There is no linear path for inputting commands, or for mapping inputs to results.
Compare this to the command line. You enter a few distinct atomic commands, and view the results in the same medium. You then enter more commands, refining your actions. The key here is that you already have a linear model for input that produces well defined expected results, all in a common medium that is conceptually simple, visible to the user, and easily processed by machines. Extending this model to accept voice input or output is trivial.
How is one supposed to quantify basic tasks and turn them into equivelant voice commands without a baseline framework or paradigm to extend from? How do you automate, simplify, or extend existing tasks without a common input or output medium? GUIs provide no such medium or framework; that same framework is at the heart of the command line interface!
Perhaps this is why we never saw voice recognition technology take off on Windows. It's blinking impossible to script actions for an arbitrary task, let alone process the arbitrary results!
On a similar note we may see voice recognition on Linux take off like a rocket. Anybody can add voice recognition to perform almost any command because the actions are all scriptable throught the CLI already. If you can type it, you can get your computer to do it when you say 'computer, foo!'
Mars
P.S. It would be greatly appreciated if someone could please clarify my point. It's buried in there somewhere...
While I've been waiting for Sphinx to mature into something useful for a long time now, the move to Java makes the whole package pretty useless to me.
:)
:P
Java is a memory hog, and it's certainly not going to be on any device I would want speech recognition on. Heck, I don't have Java installed on any of my machines, mostly because of the absolutely ridiculous footprint on disk as well as when running in ram.
And integrating Java applications into other applications is very difficult. Now, Java is good for certain things, but a speech recognition engine in Java sounds like the worst abuse possible
That and I still can't train it to recognise my slight australian accent, unlike every other bit of SR software I've used on Win32
Whether or not Sphinx-4 works, and whether or not Java is 'fast' enough to do speech recognition processing, its of no use to me.
True, instead there are a thousand "super-efficient" .jar libraries required by a "Hello World" app, which use the "Object Oriented Programing and Long Lasting Cure All and Testicular Itch Relief Paradigm(tm)" to such extremes that it takes 12 objects instantiated in 4 containers to flip a bit in a byte. Additionally, there is the substitution of native performance of compiled code to code compiled "Just Too Late" combined with exceptional memory usage that entails. If it were not enough, as a bonus, we get the garbage collector which is scientifically fine-tuned to run just when user is expected to interact with the application in most time sensitive manner. As an icing on the cake we also are treated to multiple, insideously incompatible with each other, versions of the so-called "universal" VM, resulting in one app demanding that SunVM is used and the other that MS VM is used thus resulting in total impossibility of using both at the same time. Yes I do speak from looong and utterly infuriorating experience with Java apps.
At one point I considered printing a sign warning of Java advocates being shot on sight, I could probably make some serious money selling it, given similiar amount of grief my other colleagues are going through.
Ahem, and yes, the greatest offenders in me experience are... err... frigging IBM Java apps. We actually abandoned DB2 8.x release because noone could deal with the havoc the DB2 admin tools were causing with various other retarded banking related Java apps.
Probably because speaking is incredibly complicated, and providing realistic speech from unmarked text is an intractable problem.
When you write something down, you don't provide a pronunciation guide. Rather, the reader is guided by context. For instance, if I write the word "import", how do you pronounce it? If we're talking about trade deficits, you probably know that the stress is on the second syllable; but if we're discussing meaning, the stress is on the first.
How do we expect computers that have a difficult time with context to make a pronunciation decision? This is a serious barrier to "good" text to speech (whatever "good" means).
If you mean that you want the voice to sound more natural, even if it's pronouncing words incorrectly, you still have a lot of hard problems. For instance, the muscles in the tongue and lips move differently based on how phonemes are grouped. Coarticulation models are difficult to construct, and when you try to account for a convincing number of muscles and vibrations, the problem may quickly become intractable.
Not only do we have to pay attention to the physics of speaking, but also the physics of hearing. The amount of signal processing involved can be pretty staggering if you're going to implement a complete system. Thierry Dutoit has a really good book on the subject called An Introduction to Text-to-Speech Synthesis. You should check it out if you want a somewhat more exhaustive answer to your question.
Garbage collection, in particular, is coming along nicely. Check out "Metronome" by David Bacon, of IBM. You set the knobs, it tells you how much memory you will need, and it gives you GC with real time performance. No pauses.
Or, consider the machine that Azul is working on (good luck getting details now that they are in some sort of a quiet period). It has hardware support for read and write barriers, plus a good story for stack caches. Chances are good its GC pauses will be tiny (1-10 ms).
I can also tell you that the market very much prefers JIT compilation. I worked on an ahead-of-time-compiling JVM, and there were a couple of others built by other companies. I don't work on that JVM any more, and the other AOT JVM companies have either failed or gone into other lines of business.
So, great story, but not exactly correlated with reality.
On the other hand, consider all the buggy apps that we (who sometimes administer Windows machines) have needed to patch over and over again over the years. If I am unwilling to run an application in the first place because of its poor security, does it really matter how little memory it uses, how fast it runs, or how well it gets along with the other worm-friendly apps?
I think some people should open their eyes, otherwise the world will leave you behind while you are happily consoling each other how Java is slow and unusable. Wake up, folks!
To people which argument about hand writing C and assembly - well, you obviously didn't try to implement any of the algorithms (like hidden Markov models or the statistical searches) used in speech recognition. It is pain in the butt to do it even in Java, but at least you do not have the pointer mess you would have in C/C++. The engine has a good performance already, I am not sure what you would gain by rewriting it, except of bugs (the older Sphinx2 was for sure buggy as hell).
Something about the memory footprint. Java can have a large memory footprint, however with speech recognition, you will always have it. Just the accoustic models for one language can be easily in the order of several hundreds of megabytes. Memory footprint of Java is completely irrelevant here.
And before somebody compares Sphinx with speech "recognition" on you mobile phone or in your car - be aware, that you are comparing scateboard with a Concorde here. Sphinx family of engines are intended for recognition of continuous, large vocabuly speech and to be speaker independent. Your phone/car is small vocabulary, single words and speaker dependent - i.e. completely different problem. You cannot think about Sphinx as something "to have on some device". It is more intended to act as a speech recognition server on a dedicated machine e.g. for a large call center or ticket reservation system. I guess it could be used also in KDE for the KAccessibility purposes, but it is a bit heavy for that (especially with the large datasets).
So next time, before you start spouting BS about Java and applications written in it, at least check the facts. People will not see you as a complete idiot.