envelope detection for audio transient suppression

I want to use exponential moving average filter( 1st order IIR) for
envelope detection, to apply gain for transient suppression in audio.

My requirement is that say at input x(n-1)[ when it starts rising, at onset
of transient], I compute Gain/suppression factor which is high enough to
reduce x[n-1], but by the time i compute the gain, new input x(n) would
have arrived due to filter delay, and i will be applying gain on this new
sample, rather than old sample x(n-1). so, how can I look ahead ,so to say,
that I compute gain/suppression in advance to be applied later when the
actual input sample for which it is meant arrives? what would be data
structure/buffering needed?

Thanks sunny	 

_____________________________		
Posted through www.DSPRelated.com

On 6/19/14 1:39 PM, sed_y wrote:
> I want to use exponential moving average filter( 1st order IIR) for
> envelope detection, to apply gain for transient suppression in audio.

do you mean "compression of audio"?  or "attenuation"?

>
> My requirement is that say at input x[n-1] (when it starts rising, at onset
> of transient), I compute Gain/suppression factor which is high enough to
> reduce x[n-1], but by the time i compute the gain, new input x[n] would
> have arrived due to filter delay, and i will be applying gain on this new
> sample, rather than old sample x[n-1]. so, how can I look ahead ,so to say,
> that I compute gain/suppression in advance to be applied later when the
> actual input sample for which it is meant arrives? what would be data
> structure/buffering needed?
>

"look-ahead" is done on an audio file (non-realtime) by simply looking 
ahead to samples ahead of the "present" sample.  for real time, you have 
to accept a little bit of delay.  a few samples of delay won't kill anyone.

essentially in real time, the gain factors that you compute with samples 
all the way to the most recent sample (x[n]) gets applied to some part 
of the audio a little further back in time (like x[n-D] where D is the 
delay).

-- 

r b-j                  rbj@audioimagination.com

"Imagination is more important than knowledge."

Hi Robert.

thanks for reply. I mean  "attenuation".
I have to do it real-time on DSP board.

in  Matlab, i have something like:

for j=1:N-1     % N is length of input

% x is input audio, below does 1 pole exponential averaging as difference
equation
y_fst(j+1) = (alpha1 *x(j+1)) + ((1- alpha1)* y_fst(j));  	 

_____________________________		
Posted through www.DSPRelated.com

Hi Robert.

thanks for reply. I mean  "attenuation".
I have to do it real-time on DSP board.

in  Matlab, i have something like:

for j=1:N-1     % N is length of input

    % x is input audio, below does 1 pole exponential averaging as   
difference equation
    y_fst(j+1) = (alpha1 *x(j+1)) + ((1- alpha1)* y_fst(j));  

%  some more averaging and thresholding here

     y(j+1)= x(j)*Gain_FIR_final;  % old sample


end

So, if the above computation takes more time, than my next sample arrival,
I would use  Gain_FIR_final on old sample, which should be ok?

sunny	 

_____________________________		
Posted through www.DSPRelated.com

On 6/19/14 3:38 PM, sed_y wrote:
> Hi Robert.
>
> thanks for reply. I mean  "attenuation".
> I have to do it real-time on DSP board.
>
> in  Matlab, i have something like:
>
> for j=1:N-1     % N is length of input
>
>      % x is input audio, below does 1 pole exponential averaging as
> difference equation
>      y_fst(j+1) = (alpha1 * x(j+1)) + ((1- alpha1)* y_fst(j));

what's the x(j+1) for?  are you intending to skip x(1)?

are you sure you don't want |x(j)| or |x(j)|^2 going into the weighted 
moving average?

>
> %  some more averaging and thresholding here
>

how 'bout

         Gain_FIR_final = somefunction(y_fst(j-D));

(for the case where j-D>0) ?

and where D = something like 4 or 5 time-constants (i think it's 
1/alpha1 in sample units).

>       y(j+1)= x(j)*Gain_FIR_final;  % old sample
>

and
         y(j) = x(j-D)*Gain_FIR_final

>
> end
>
> So, if the above computation takes more time, than my next sample arrival,
> I would use  Gain_FIR_final on old sample, which should be ok?
>
> sunny	
>
> _____________________________		
> Posted through www.DSPRelated.com


-- 

r b-j                  rbj@audioimagination.com

"Imagination is more important than knowledge."

>On 6/19/14 3:38 PM, sed_y wrote:

>what's the x(j+1) for?  are you intending to skip x(1)?
>
>are you sure you don't want |x(j)| or |x(j)|^2 going into the weighted 
>moving average?

Hi Robert,
No I am not intending to skip x(1)[ I was evaluating it outside the loop 
but did not put it there ].

I am actually sending in abs(x) as the input.



>> %  some more averaging and thresholding here
>>
>
>how 'bout
>
>         Gain_FIR_final = somefunction(y_fst(j-D));
>
>(for the case where j-D>0) ?
>
>and where D = something like 4 or 5 time-constants (i think it's 
>1/alpha1 in sample units).
>
>>       y(j+1)= x(j)*Gain_FIR_final;  % old sample
>>
>
>and
>         y(j) = x(j-D)*Gain_FIR_final
>
>>
>> end

suppose j=30
with time constant of 6 samples, 4 time constants =24 =D
so, within loop at time j=30 sample
i will have 

Gain_FIR_final = somefunction(y_fst(6));
y(30) = x(6)*Gain_FIR_final

if the transient starting appearing at j=10, I would be applying
attenuation computed at the high input sample at j=30, to the sample at
j=6, which was not a transient and so, should not be applied this high
attenuation.?

sedy	 

_____________________________		
Posted through www.DSPRelated.com

please read j as 30 in the previous post.
pleas read D as 24 in the previous post.	 

_____________________________		
Posted through www.DSPRelated.com

On 6/20/14 12:24 PM, sed_y wrote:
>> On 6/19/14 3:38 PM, sed_y wrote:
>
>> what's the x(j+1) for?  are you intending to skip x(1)?
>>
>> are you sure you don't want |x(j)| or |x(j)|^2 going into the weighted
>> moving average?
>
> Hi Robert,
> No I am not intending to skip x(1)[ I was evaluating it outside the loop
> but did not put it there ].
>
> I am actually sending in abs(x) as the input.
>

okay, you have mean abs, rather than just mean.  that's okay.  dunno 
which is better.  sometimes you want *peak* abs value (over a finite 
period behind the "current" sample.

okay, i'm not gonna write code for you (except maybe a line here or 
there) and i am not gonna write code to test in MATLAB (or Octave).

the issue you're having is that of "edge conditions" (which i would not 
conflate with "boundary conditions" involving differential equations or 
difference equations - essentially that of setting up the states to 
initial conditions before you start processing samples).

my recommendation for dealing with these edge effects is to zero-pad 
your data on both sides.  and then start with sample #1 (or sample #0, 
if you were doing this in a language that can have 0 origin) of the 
original data, which would be sample N+1 if you pad with N zeros before 
the first sample.

so you know how to concatenate arrays in MATLAB, so then concatenate N 
zeros before your 1st sample (which makes it the (N+1)th sample in the 
new concatenated array).  then start processing from that sample, in 
computing your mean abs.  but apply the results of that to the sample D 
samples preceding (and D better not exceed N).  this is how you would 
implement this in C or assembly or some non-MATLAB setting.


-- 

r b-j                  rbj@audioimagination.com

"Imagination is more important than knowledge."

Thanks a lot	 

_____________________________		
Posted through www.DSPRelated.com

On Thursday, June 19, 2014 1:39:49 PM UTC-4, sed_y wrote:
> I want to use exponential moving average filter( 1st order IIR) for
> 
> envelope detection, to apply gain for transient suppression in audio.
> 
> 
> 
> My requirement is that say at input x(n-1)[ when it starts rising, at onset
> 
> of transient], I compute Gain/suppression factor which is high enough to
> 
> reduce x[n-1], but by the time i compute the gain, new input x(n) would
> 
> have arrived due to filter delay, and i will be applying gain on this new
> 
> sample, rather than old sample x(n-1). so, how can I look ahead ,so to say,
> 
> that I compute gain/suppression in advance to be applied later when the
> 
> actual input sample for which it is meant arrives? what would be data
> 
> structure/buffering needed?
> 
> 
> 
> Thanks sunny	 
> 
> 
> 
> _____________________________		
> 
> Posted through www.DSPRelated.com

The way to predict without introducing lag is to use derivatives to project forward in time.   
Look at alpha/beta filters for some low complexity ideas.