Sign in

Not a member? | Forgot your Password?

Search compdsp

Search tips


Discussion Groups

FFT Spectral Analysis Software

Free Online Books

See Also

Embedded SystemsFPGA

Discussion Groups | Comp.DSP | Unsupervised Spectral Subtraction?

There are 4 messages in this thread.

You are currently looking at messages 1 to .


Is this discussion worth a thumbs up?

0

Unsupervised Spectral Subtraction? - Steve - 2006-02-04 00:35:00

Hello Everybody,
I have posted here before, but now I know a little more about what I want to 
do.

I left most of my math skills back at the college door, so when I look at 
some of these dsp algorithms, my head spins. However, I am now in need of a 
java implementation of some sort of unsupervised noise removal.

Specifically, I need something that will remove the static from a recording 
of a police scanner. The static is very often different, so there can't be 
any hard coding of the noise estimation. Basically, I need to perform some 
speech recognition (looking for some key words) on the recordings, but the 
static is getting in my way.

1. Is spectral subtraction the way to go?
2. Are there some java implementations out there for me to use?
3. If you can do this, I may be willing to pay. Of course, it would need to 
perform the functionality (ie the static must be removed enough to perform 
the simple speech recognition that am talking about on MULITPLE samples).
I can't pay much... this is just an idea that I am working on. If you can 
help for free, great. If not, drop me an email and let me know what you 
would charge. I am but a poor dude with just an idea and very little
budget.


Re: Unsupervised Spectral Subtraction? - Bob the Builder - 2006-02-04 04:09:00

"Steve" <s...@saddleback.edu> wrote in message
news:D...@comcast.com...
> Hello Everybody,
> I have posted here before, but now I know a little more about what I want
to
> do.
>
> I left most of my math skills back at the college door, so when I look at
> some of these dsp algorithms, my head spins. However, I am now in need of
a
> java implementation of some sort of unsupervised noise removal.
>
> Specifically, I need something that will remove the static from a
recording
> of a police scanner. The static is very often different, so there can't be
> any hard coding of the noise estimation. Basically, I need to perform some
> speech recognition (looking for some key words) on the recordings, but the
> static is getting in my way.
>
> 1. Is spectral subtraction the way to go?
> 2. Are there some java implementations out there for me to use?
> 3. If you can do this, I may be willing to pay. Of course, it would need
to
> perform the functionality (ie the static must be removed enough to perform
> the simple speech recognition that am talking about on MULITPLE samples).
> I can't pay much... this is just an idea that I am working on. If you can
> help for free, great. If not, drop me an email and let me know what you
> would charge. I am but a poor dude with just an idea and very little
budget.
>
>
>

With narrow band FM which is police radio - the quality is not too good for
any kind of recognition.
Also, the speech recognition algorithms are only good when they are trained
for a given voice.
For un-trained they can be ok but on a limited vocab -so you might be ok
there. It might work but what hit rate I have no idea.


Bob


Re: Unsupervised Spectral Subtraction? - Steve - 2006-02-04 09:46:00

Thanks for the response bob...
I am only looking for a couple keywords that have relatively obvious 
phonemes. Plus, I can make the SR more sensitive, so that it will hit more 
false positives, this helps. Plus, for me to catch what I am looking for, 
the keywords are mentioned multiple times in multiple transmissions, so I 
have plenty of opportunity to catch them.

As for the SR and training, I am using Microsoft's SAPI 5, it seems to pick 
off the keywords when I use a microphone and when my wife uses it. I believe 
training may or may not be an issue.

Any ideas on algorithms?

"Bob the Builder" <F...@yahoo.co.zpr> wrote in message 
news:ib_Ef.129200$v...@news.xtra.co.nz...
>
> "Steve" <s...@saddleback.edu> wrote in message
> news:D...@comcast.com...
>> Hello Everybody,
>> I have posted here before, but now I know a little more about what I
want
> to
>> do.
>>
>> I left most of my math skills back at the college door, so when I look
at
>> some of these dsp algorithms, my head spins. However, I am now in need
of
> a
>> java implementation of some sort of unsupervised noise removal.
>>
>> Specifically, I need something that will remove the static from a
> recording
>> of a police scanner. The static is very often different, so there can't

>> be
>> any hard coding of the noise estimation. Basically, I need to perform 
>> some
>> speech recognition (looking for some key words) on the recordings, but

>> the
>> static is getting in my way.
>>
>> 1. Is spectral subtraction the way to go?
>> 2. Are there some java implementations out there for me to use?
>> 3. If you can do this, I may be willing to pay. Of course, it would
need
> to
>> perform the functionality (ie the static must be removed enough to 
>> perform
>> the simple speech recognition that am talking about on MULITPLE
samples).
>> I can't pay much... this is just an idea that I am working on. If you
can
>> help for free, great. If not, drop me an email and let me know what
you
>> would charge. I am but a poor dude with just an idea and very little
> budget.
>>
>>
>>
>
> With narrow band FM which is police radio - the quality is not too good 
> for
> any kind of recognition.
> Also, the speech recognition algorithms are only good when they are 
> trained
> for a given voice.
> For un-trained they can be ok but on a limited vocab -so you might be ok
> there. It might work but what hit rate I have no idea.
>
>
> Bob
>
>


Re: Unsupervised Spectral Subtraction? - Bob the Builder - 2006-02-04 22:22:00

"Steve" <s...@saddleback.edu> wrote in message
news:N...@comcast.com...
> Thanks for the response bob...
> I am only looking for a couple keywords that have relatively obvious
> phonemes. Plus, I can make the SR more sensitive, so that it will hit more
> false positives, this helps. Plus, for me to catch what I am looking for,
> the keywords are mentioned multiple times in multiple transmissions, so I
> have plenty of opportunity to catch them.
>
> As for the SR and training, I am using Microsoft's SAPI 5, it seems to
pick
> off the keywords when I use a microphone and when my wife uses it. I
believe
> training may or may not be an issue.
>
> Any ideas on algorithms?
>

Yes the Microsoft engine is good for independant word recognition-  I woudl
use that. As for noise reduction - start by band-limiting the signal.
Spectral subtraction only works well for stationary type noises and it
suffers from musical noise though there are quite sophisticated algorithms
to minimise musical noise. I don't hold out much hope - you woudl need to
make a recording and do some off-lien tests and see what you are dealing
with.


Bob