Registering and logging in removes this ad.
Registering and logging in removes this ad.
Dragon on the cheap, part 2: the $9 sound card
Submitted by DGehman on Sun, 08/13/2006 - 11:16.
[Dragon on the cheap is a series of articles about how to achieve just about the highest level of voice recognition, using high-quality, low-cost equipment and approaches. This article features a $9 USB sound adapter from, no surprise, eBay.]
The second most important element in good speech recognition is a decent sound card. (The first element is your diction, how you speak your computer.)
Most computers today are supplied with onboard sound chips, and these are notoriously bad for voice recognition. Onboard audio is nestled in with a bunch of electronically noisy computer components.
Some people cough up $150-$400 for a professional audio card, not knowing that voice recognition requires not high Fidelity, but responsiveness in a fairly narrow audio band. Others, myself included, pop for the USB sound "pods" designed for voice recognition, such as the $40 Andrea.
I have been more than happy with the Andrea, except for an incredibly cheap microphone jack that broke. But, on a whim, and in pursuit of my unwavering goal to do speech recognition at a laughably low cost, I turned to that endless fount of cheap Chinese components, eBay.
After extensive research, I opted for one that cost just under $9, including shipping. The research consisted of finding (a) the lowest price including shipping, and (b) anything not shipped from Hong Kong.
The little unit works. It works flawlessly in version 7 .3 but has an odd glitch in my brand new version 8 -- after starting NaturallySpeaking, I have to unplug the thing, then plug it back in again, before NatSpeak is able to listen to it.
****************************************
UPDATE August 14: The problem was that I routed this adapter through a USB 1.1 hub. Plugged directly into a USB port, NaturallySpeaking sees this sound adapter on start-up. /END UPDATE
****************************************
I also had to shave a flat spot on my Andrea headset's microphone and earpiece plugs because the two input jacks on the sound adapter are a tad too close together.
On my system, it is a little noisier than the Andrea (the $9 adapter has a noise floor of -41dB, where the Andrea's noise floor is -60dB, measurably better), but I cannot hear the difference in playback at ordinary volume, and apparently neither can NaturallySpeaking.
It is based on the C-Media CM-108 combination audio/USB chip. Technically, this chip cannot operate at the 11.025 kHz sampling rate required by NaturallySpeaking. But, like the bumblebee who flies despite the engineering calculations that prove it can't, it works well with NaturallySpeaking.
It comes in a nifty dark blue, quasi-wedge shaped case. Installation required only plugging into Windows XP, which dutifully found its C-Media driver somewhere in the XP innards.
The adapter features an LED that glows when it is plugged in and recognized by the computer, and blinks when you turn on the microphone in NaturallySpeaking. It seems to be saying with each blink, "talk to me, talk to me, talk to me..."
I do indeed talk to it. I have to admit, I am sometimes distracted by thinking about the cash that I'm saving, but other than that, it works just fine.
»
- DGehman's blog
- Login or register to post comments
Dave wrote: , After
,
After extensive research, I opted for one that cost just under $9, including shipping. The research consisted of finding (a) the lowest price including shipping, and (b) anything not shipped from Hong Kong.
The little unit works. It works flawlessly in version 7 .3 but has an odd glitch in my brand new version 8 -- after starting NaturallySpeaking, I have to unplug the thing, then plug it back in again, before NatSpeak is able to listen to it. I also had to shave a flat spot on my Andrea headset's microphone and earpiece plugs because the two input jacks on the sound adapter are a tad too close together. On my system, it is a little noisier than the Andrea (the $9 adapter has a noise floor of -41dB, where the Andrea's noise floor is -60dB, measurably better), but I cannot hear the difference in playback at ordinary volume, and apparently neither can NaturallySpeaking.I do indeed talk to it. I have to admit, I am sometimes distracted by thinking about the cash that I'm saving, but other than that, it works just fine.
It is obvious that the $40 USB sound pods designed for speech recognition are built to better tolerances and specifications as well as conforming to the audio standards of Microsoft Windows so you do not have to unplug/replug the device each time you start NaturallySpeaking. I think most people would prefer to have the real thing. However, I can certainly understand how a computerphile like you is willing to play and discover.
Marty
Martin Markoe wrote: It is
It is obvious that the $40 USB sound pods designed for speech recognition are built to better tolerances and specifications as well as conforming to the audio standards of Microsoft Windows so you do not have to unplug/replug the device each time you start NaturallySpeaking. I think most people would prefer to have the real thing. However, I can certainly understand how a computerphile like you is willing to play and discover.
Marty
Sad to say, the components (USB connector, mic & ear jacks), soldering and overall efficiency of layout are better in the $9 unit than in my Andrea. But, I bought my Andrea unit years ago, so perhaps they have gone up in quality. The audio performance of the Andrea as reflected in CoolEdit is measurably better, as I said.
And I should point out that there is no engineering to speak of in the $9 unit. It simply replicates C-Media's reference circuit, even to the point of including up/down volume and mute switch pads (unpopulated), and for half of the board, using C-Media's numbers for the devices, capacitors and resistors.
As for the usability, I also have to plug and unplug my Epson USB printer, and used to have to do that same with my USB mouse, so the jury is out on whether my computer's USB is weird, or the audio unit is not up to Windows standards. The mouse finally settled down.
UPDATE:
The USB problem was that I was trying to connect via a USB hub. There is no trouble with the $9 adapter, the printer or the mouse, when plugged directly into a USB port on my computer.
/END UPDATE
And "computerphile"? Yeah, I guess. I was born the same year that Atanasoff and Berry in Iowa built the first all-electronic digital computer (1942), and I'm still totally fascinated by them -- computers, that is. I never met Atanasoff & Berry, so I have no idea how fascinating they were.
But I'm really just a cheapo-phile. My 1990 Taurus was free (and worth every penny), my FM tuner/amp is a Miita (who?) from 1976, and I recently added an extra 9GB 10,000 rpm SCSI drive to my computer for $5 + gas (another $5).
Dave
"Wait a minute," I says to myself later, waiting in line at a warehouse club to buy hot dogs at 1/10 the price of the supermarket, "you've forgotten customer service."
Yep. I don't have the smarts or the equipment to do extensive testing. I don't know if the next eBay sound adapter is going to be as good as mine, or pure garbage. And nobody is going to stand behind the unit.
So Marty's right, if you're serious, you're likely to want to go for 'the real thing.'
(Not) Extensive Testing... I
(Not) Extensive Testing...
I decided to put the nine dollar audio adapter to the test -- that is, by reading the Rainbow Passage test that has become a sort of de facto touchstone.*
I dictated it into my brand new NaturallySpeaking version 8, using the much denigrated headset that came with my version 6 NaturallySpeaking, the Andrea NC-61. All over the world, time and time again, people have been counseled to throw this away and buy a real microphone.
I did no training of NaturallySpeaking vis-à-vis this passage. I simply read it straight in, once, and let the cards fall where they may.
They fell pretty well: six errors in 335 words for 98.2% accuracy.
In my estimation, four of the errors are forgivable. One further one happened to me no matter how many times I read it in to NaturallySpeaking version 7. Only one extremely puzzling error might suggest some sort of noise or electronic glitch.
The errors:
"past" for "path" (forgivable)
"birth" for "Earth' (one that repeats for me, probably because the context is, "over which the gods passed from Earth to their home" and I sort of get a lip bounce when I am pronouncing final "m"s).
"phenomenon" for "phenomena" (forgivable)
"color" for "colored" (forgivable)
"line at" for "band" (mighty strange...)
"bowl" for "bow" (forgivable)
On sober reflection, "color" for "colored" is perhaps not forgivable, as it's one of those -ED recognition errors that drive you nuts in NaturallySpeaking.
I did not try reading the passage using my primo handheld speech recognition microphone, the Shure VR-116L that I luckily bought before Shure stopped selling it.
Dave
*This passage can be downloaded on Martin Markoe's site, http://www.emicrophones.com/microphones/howtest.as...
Here, you can also read about the far more extensive test suite used by Martin for real microphones and real audio adapters.
If so, then this is a fairly
If so, then this is a fairly poor showing for the $9 adapter because the Rainbow passage no touchstone at all but rather a creampuff for an SR recognizer. The first time I read it, I expect maybe one or two errors at most if I'm having a bad day -- in other words, I expect no errors. (Each repetition will produce successively more errors because of muscular fatigue.)
BTW, its an easy test because it supposedly contains all the phonemes in the English language, giving a recognizer a fairly distinctive target -- much harder is a routine bit of text with lots of exact and near homonyms.
BTW #2, I don't get the "forgivable" shtik at all -- an error is an error. When people start classifying their errors I think its a PR ploy for special pleading, i.e., an argument that can't stand on its own. What's forgivable for you might be an unpardonable sin for me, and vice-versa
Call them all unqualified errors and be done with it.
However, in this context a one-legged test has no face cred -- it needs to be compared to a test using your best input sound system (but still the same microphone if your purpose is to demonstrate the capability of the soundpod per se).
FWIW, I found your piece on buying a used server more interesting and relevant to my needs -- which is not to say I don't think you shouldn't continue this piece because it might be useful for others.
Bruce
DGehman wrote: I decided to
I decided to put the nine dollar audio adapter to the test -- that is, by reading the Rainbow Passage test that has become a sort of de facto touchstone.*
What was the ambient noise like when you did your test?
Martin
Martin Markoe wrote: What
What was the ambient noise like when you did your test?
Martin
Basically, quiet. White noise from an old-style HEPA air filter (3 dB quieter when off, using CoolEdit & Shure VR-166L.) My computer is very quiet. Some aircon noise (less than 1dB per above so call "instrumentation"). Some intermittent road noise (my picture window overlooks the Mass Turnpike).
Dave