Hmmm... Whose voice-recognition software is it?

In past announcements, the Empeg folks have referred to the voice-recognition software as being "licensed from a well-known Japanese company" or something like that. Otherwise, they've been mum about it.

So I'm reading this back issue of PC magazine (Jan 18th 2000), and there's a blurb on page 11 about a wrist-phone with voice recognition. Supposedly, you'll be able to compose e-mail on this thing. Hmmm... let's look at the particulars of this little wrist-phone:

- Runs on a 30mhz ARM processor.
- Voice recognition designed to work in a very small RAM footprint.
- Wrist phone is made by Samsung.

So I'm thinking, "hey, I'll bet this uses the same software". But this wrist phone licenses its software from Conversa, which is a US-based company, not a Japanese company. Darn. I thought I had a scoop.

Okay, Rob, spill it. Whose voice recognition software are you using?



http://Tony Fabris
Empeg #144

I'm not sure if we can say at the moment, I'll check and get back to you.

Rob

Did some search and found a candidate:

http://www.oki.co.jp/semi/english/t-vrp.htm

They are a StrongARM integrator...

Gary

The recogition chip specs indicate it will misinterpret one out of every ten words. In trained mode it misses ~half as many. It has a maximum vocabulary of 61 words. It requires a signal to noise ratio of >20db.

Calvin

We're not using an external chip, so ignore the limitations of the OKI part - ours is software only (well, plus the sound input hardware!)

Hugo

What will be the ramifications of non-english-speaking users and voice recognition on the empeg? Will it be multi-lingual? Will we be able to program (i.e., "teach") it to do anything we want? Could one, theoretically, say "'Twas brillig and the slithy toves" and have it select equalization setting #4 and set the volume to +5dB? And just think of the possibilities inherent in mome raths and frumious bandersnatches!

Hmmm... how about some sort of security centered around voice recognition: the unit won't turn on unless it recognizes MY voice saying "Good morning, empeg." Will the voice recognition control the empeg's external signals -- can I say "Empeg -- antenna up" and have the antenna go up? That would surprise and amaze a few people, wouldn't it. (Especially on my car which doesn't have a retractable antenna!)

What would be really great is if you or Rob could make a major posting to this bbs describing all the things you have implemented, plan to implement, or maybe someday might think about implementing pertaining to voice control on the empeg.

tanstaafl.

"There Ain't No Such Thing As A Free Lunch"

As I understand it (reading from previous posts), it'll have English recognition of a few specific phrases that are speaker-independent (i.e., you don't have to train it, but they would be language-dependent), and then everything after that will be trainable (and therefore will be speaker-dependent as well as language-independent).

I don't know how detailed the trainable functions will be. The idea of a single command having multiple results would require a macro facility to be built into the Empeg: something that doesn't exist yet as far as I know. I'm sure Rob or Hugo will jump in here and give some hints, but I wouldn't be surprised if they declined to publish any details yet. They're still working on it and they probably won't want to promise features that don't exist yet.

Now, as far as using voice for security goes, I don't see that happening. The issues about security have already been hashed through in other forums.



Tony Fabris
Empeg #144

We still have a lot of work to do on the voice recognition software (which is mainly John's pet project) so the in-depth description will have to wait for a while. I would guess that voice recognition will be considered a beta test element of the new player when it first ships (maybe there will be a "safe" consumer release that doesn't incorporate it right away).

The multi-lingual element is something I was just thinking about a few moments ago. Some elements of the system are speaker independent, so it would be necessary for us to either provide alternative nets for different languages or to offer the option to make them speaker dependent for non-english speaking clients (and for those in the deep south of the USA - joke).

If we resume this conversation in a month or so I would guess we'll be able to fill in a lot more detail then.

Rob

Will it be multi-lingual? (...) Could one, theoretically, say "'Twas brillig and the slithy toves"

You mean "To bje u przisat, i otke tajke..." ?

Dragi "Bonzi" Raos
Zagreb, Croatia
#5196