Audio: TDFB: Track sound direction of arrival by singalsu · Pull Request #4826 · thesofproject/sof

singalsu · 2021-09-29T14:13:06Z

This patch adds sound direction tracking. TDFB component updates
the sound direction control to user space. User space may control
the beam direction with help of this.

Signed-off-by: Seppo Ingalsuo seppo.ingalsuo@linux.intel.com

singalsu · 2021-09-29T14:17:40Z

First draft, needs a lot of cleanup and testing and optimization. It tracked in simulated 2mic acoustics in testbench quite well a 360 degree rotating sound source within -90 .. +90 deg that the 2D array can handle.

singalsu · 2021-10-04T16:45:20Z

I just pushed a version that "worked" correctly in a notebook device. I will next tune it to be more responsive, now only fairly loud sounds trigger angle update. Also I should change the enum scale for line array topologies for better precision. Alsatplg allows 16 different values and now 12 is used to cover 0-330 degrees. I could use 15 steps for -90..0..+90 with ~13 degrees precision. Or 13 steps with exactly 15 degrees step.

lgirdwood · 2021-10-15T14:54:48Z

I just pushed a version that "worked" correctly in a notebook device. I will next tune it to be more responsive, now only fairly loud sounds trigger angle update. Also I should change the enum scale for line array topologies for better precision. Alsatplg allows 16 different values and now 12 is used to cover 0-330 degrees. I could use 15 steps for -90..0..+90 with ~13 degrees precision. Or 13 steps with exactly 15 degrees step.

Best to use integer step size for alsamixer rendering.

singalsu · 2021-10-18T08:09:57Z

Best to use integer step size for alsamixer rendering.

Yep, my working version is using exact 15 degree step for line arrays now. I will share soon a testbench PR to enable testing of different input and output channels count in pipeline and improved TDFB test scripts. Then share the update to this PR.

singalsu · 2021-11-17T16:34:07Z

This PR needs #4941. I will meanwhile share related parts to this work (TDFB comp bug fixes, TDFB tool improvement, testbench improvement, TDFB test improvement). Also avoids huge single PR since I had to change plenty of things to be able to test this. This version with internal math functions remains [WIP].

lgirdwood · 2021-11-18T16:30:16Z

@singalsu ok pls work with @ShriramShastry for #4941. Changes are localized so can be included for rc2

lgirdwood · 2021-11-23T13:30:20Z

@singalsu ready for review now ?

singalsu · 2021-11-26T17:02:32Z

@singalsu ready for review now ?

Sorry, not yet. Need to solve issue in #4941 to compute square root, the power function there is not suitable for that. There's a another PR coming for square root. Backup is to convert power function part of DRC to math library.

singalsu · 2021-11-26T18:06:39Z

@lgirdwood I'm now testing this with PR #5034 .

singalsu · 2021-11-30T16:46:36Z

New WIP draft with Sriram's sqrt() lib patch temporarily included.

lyakh · 2021-12-22T08:23:20Z

is this the same as ((a + 180) % 360) - 180? Maybe the below can be similarly simplified too, but be careful with division remainder of negative numbers

Thanks, that worked, and slightly different version for negative angles.

lyakh · 2021-12-22T08:51:31Z

nitpicking: how about
cd->direction.rp = cd->direction.d; cd->direction.wp = cd->direction.d + ch_count * (cd->direction.max_lag + 1);

Good catch, this was a quick bug fix test that I forgot to clean up.

lyakh · 2021-12-22T08:53:08Z

you could use goto for clean up

lyakh · 2021-12-22T08:55:59Z

cd->direction.trigger <<= 1;

lyakh · 2021-12-22T09:11:17Z

something above makes me suspicious. If the first test ds2 < ds1 was true, 360 was subtracted from az. And now you add 360 back to az for another test? An else missing somewhere?

There could be an else. The second squared error eval is unnecessary if the first 360 deg addition was done. The second test is false if first was true. One multiplication saved in that case.

lyakh · 2021-12-22T09:26:25Z

unsigned int?

OK, changed!

lgirdwood · 2021-12-23T20:34:24Z

@cujomalainey good for you ?

singalsu · 2022-01-18T15:15:37Z

I just updated the emphasis IIR response from 2nd order 2 kHz high-pass to 1 - 4 kHz 4th order band-pass. It prevents highest frequencies to alias spatially and add error to angle. The lower high-pass picks up a bit more voice signal while still attenuates lowest frequencies those are usually noise. No other changes.

lgirdwood · 2022-01-26T11:04:10Z

@singalsu any updates ?

The DF2T IIR core has been moved earlier to library but the helper functions for configuring it were left to EQ component. This patch eases use of IIR core from other components. The file src/audio/eq_iir/iir.c is moved to src/math. There are no changes except of remove iir.h since the contents specific for DF2T type were merged to iir_df2t.h Signed-off-by: Seppo Ingalsuo <seppo.ingalsuo@linux.intel.com>

This patch adds function cross_product_s16(). Cross product is commonly used function in computational geometry. Signed-off-by: Seppo Ingalsuo <seppo.ingalsuo@linux.intel.com>

This patch adds sound direction estimation. TDFB component updates the sound direction control to user space. User space may control the beam direction towards the source with help of the reported angle. Sound direction estimation is based on cross correlation maximums pattern and match of theoretical propagation delays to measured pattern. Cross correlation search is triggered by short term level when it exceeds the estimated background noise level. An emphasis filter is used to filter out typical low frequency noise. Signed-off-by: Seppo Ingalsuo <seppo.ingalsuo@linux.intel.com>

Comment text is added to tdfb.c to help find the settings generator shell script. Signed-off-by: Seppo Ingalsuo <seppo.ingalsuo@linux.intel.com>

singalsu · 2022-02-02T15:51:06Z

I just added a commit that adds to tdfb.c a comment about settings generator script to help find it. No other changes done.

lgirdwood · 2022-02-03T16:13:29Z

@cujomalainey good for you ?