Low-pass-colored output from filter-and-sum beamformer

Hi all,

I'm currently trying to apply some front-end processing to microphone
signals before given the processed signal to an an automatic speech
recognizer.  Of course, the purpose of pre-processing is to enhance the
signal, i.e., to remove unwanted interference from the recorded microphone
signals.

One of the things I'm doing is beamforming. I'm using a filter-and-sum
beamformer. The corresponding beampattern shows that there is a decreasing
beamwidth as the frequency increases, so at very low frequencies the main
beam is very wide and it narrows towards higher frequencies. When I listen
to the beamformer's output signal, it sounds like that low frequencies are
amplified relative to higher frequencies. This is especially disturbing if
there are some low-frequency interferers in the recording like a bus
passing by.

Do you have any hints/suggestions what to do about this? I can't change
the array geometry.
My first (maybe naive) approach was to suppress everything below 300Hz (to
avoid amplification of low-frequency interference), but then of course,
the output signal does not sound natural.

I would be very thankful if you could help me here :-)



---------------------------------------
Posted through http://www.DSPRelated.com

> 
> Do you have any hints/suggestions what to do about this? I can't change
> the array geometry.
> My first (maybe naive) approach was to suppress everything below 300Hz (to
> avoid amplification of low-frequency interference), but then of course,
> the output signal does not sound natural.
> 
> I would be very thankful if you could help me here :-)
> 

So try 100Hz.
Low cut everything below 100Hz, this is routine for processing voice (not music) signals.

Next, see if adding a  +6 dB per octave EQ between 100 Hz and 5 kHz sounds better. 

On 6/19/15 1:37 PM, makolber@yahoo.com wrote:
>
>>
>> Do you have any hints/suggestions what to do about this? I can't change
>> the array geometry.
>> My first (maybe naive) approach was to suppress everything below 300Hz (to
>> avoid amplification of low-frequency interference), but then of course,
>> the output signal does not sound natural.
>>
>> I would be very thankful if you could help me here :-)
>>
>
> So try 100Hz.
> Low cut everything below 100Hz, this is routine for processing voice (not music) signals.

what happens for people sitting right in the center of the beam?  won't 
they get an low cut version of the audio?

i've always thunk that beamforming for audio was a tougher problem than 
the corresponding beamforming done for RF (with phased antennae) or even 
acoustical beamforming done with ultrasonic because of the bandwidth in 
octaves (or the Q).  as long as the delta_f was small, then the 
delta_wavelength would be small and the math for the low frequencies 
comes out to be about the same as for the high frequencies allowed 
within the narrow passband.

but in audio, you could have several octaves of bandwidth and it just 
ain't gonna come out the same off axis as it does on axis.  so if you 
try to correct the off-axis behavior (say with a low-cut filter), you'll 
"correct" it on-axis as well.

seems to me that there would have to be quite different filtering done 
on the transducers on the side vs. those in the middle to fix this.  the 
spacing between transducers cannot be changed as a function of 
frequency, but the gain and phase adjustment for the transducers on the 
side can be different at low frequencies than the gain and phase 
adjustment for the transducers in the middle.

i'm probably being pedantic (my favorite word of the day).


-- 

r b-j                  rbj@audioimagination.com

"Imagination is more important than knowledge."

On Saturday, June 20, 2015 at 1:44:44 AM UTC+12, henne wrote:
> Hi all,
> 
> I'm currently trying to apply some front-end processing to microphone
> signals before given the processed signal to an an automatic speech
> recognizer.  Of course, the purpose of pre-processing is to enhance the
> signal, i.e., to remove unwanted interference from the recorded microphone
> signals.
> 
> One of the things I'm doing is beamforming. I'm using a filter-and-sum
> beamformer. The corresponding beampattern shows that there is a decreasing
> beamwidth as the frequency increases, so at very low frequencies the main
> beam is very wide and it narrows towards higher frequencies. When I listen
> to the beamformer's output signal, it sounds like that low frequencies are
> amplified relative to higher frequencies. This is especially disturbing if
> there are some low-frequency interferers in the recording like a bus
> passing by.
> 
> Do you have any hints/suggestions what to do about this? I can't change
> the array geometry.
> My first (maybe naive) approach was to suppress everything below 300Hz (to
> avoid amplification of low-frequency interference), but then of course,
> the output signal does not sound natural.
> 
> I would be very thankful if you could help me here :-)
> 
> 
> 
> ---------------------------------------
> Posted through http://www.DSPRelated.com

as a matter of interest, how are you designing this beamformer. Do you have a paper or method in particular?

>as a matter of interest, how are you designing this beamformer. Do you
have
>a paper or method in particular?

Hi,
for the beamformer design I try to approximate a desired response in the
least-squares sense subject to a distortionless constraint in look
direction and a constraint on the whinte noise gain. I found this design
in the following paper:
http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=4959524&queryText=mabande&newsearch=true&searchField=Search_All

I tried lowering the cut-off frequency from 300Hz to 100Hz. That makes the
output signal sound better/more natural (which can be expected I guess)
but leads to the situation that low-frequency noise cannot be suppressed
that well anymore. The beamformer itself can only perform some spatial
filtering above 500Hz due to the array geometry.
---------------------------------------
Posted through http://www.DSPRelated.com

On Tue, 23 Jun 2015 06:27:44 -0500, "henne"
<50716@DSPRelated> wrote:

>>as a matter of interest, how are you designing this beamformer. Do you
>have
>>a paper or method in particular?
>
>Hi,
>for the beamformer design I try to approximate a desired response in the
>least-squares sense subject to a distortionless constraint in look
>direction and a constraint on the whinte noise gain. I found this design
>in the following paper:
>http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=4959524&queryText=mabande&newsearch=true&searchField=Search_All
>
>I tried lowering the cut-off frequency from 300Hz to 100Hz. That makes the
>output signal sound better/more natural (which can be expected I guess)
>but leads to the situation that low-frequency noise cannot be suppressed
>that well anymore. The beamformer itself can only perform some spatial
>filtering above 500Hz due to the array geometry.

I've never dabbled with speech recognition methodology, but
note that 300-3000 Hz is a long-standing tradition for
"telephone quality" speech.  There isn't really much
information below 300 Hz, just male fundamental
glottal-pulse frequencies.  (Nobody seems to miss them over
the telephone.)  It's the higher regions with formants, etc,
that have all the speech info.

Best regards,


Bob Masta
 
              DAQARTA  v8.00
   Data AcQuisition And Real-Time Analysis
              www.daqarta.com
Scope, Spectrum, Spectrogram, Sound Level Meter
 Frequency Counter, Pitch Track, Pitch-to-MIDI 
FREE 8=channel Signal Generator, DaqMusiq generator    
          Science with your sound card!