How it works

How the correction works.

The chest wall sits between the heart and every microphone you can put on the skin. The insert estimates its transmission from the recording and divides the signal by it. No trained model and no second sensor.

The correction

Ĥ(f) = H(f) / T(f, d̂)

H(f) is the recorded signal, T is the estimated transmission of the chest wall, and d̂ is the wall thickness estimated from the same take. Every term can be checked from the recording.

What happens in six seconds.

Listen

One channel, 20 Hz to 2 kHz. The reference capsule sits behind the face gasket and hears the room.

Measure

The low frequency slope of the take gives the wall thickness. RMS error is 3.7 mm across 111 recordings.

Divide

The murmur band is divided by the estimated transmission. It is arithmetic, not a learned correction.

Decide

Contact above 45 per cent keeps the take. Below it, the app returns no result.

Check the method yourself.

The bench recordings, the estimator and the spreadsheet behind each number are available on request, under NDA if the CAD is included.

Reproduce the spreadtwo recordings

Take the two bench recordings of the same synthetic murmur through 60 mm and 110 mm walls. Measure the murmur band level in each. That difference is the 1.73 dB. Apply the correction. That is the 0.25 dB.

Check the estimator111 recordings

The wall estimate is a slope fit over 20 to 120 Hz. It has no trained weights and no lookup table: given the same recording it returns the same number, and the residual against the measured wall is 3.7 mm RMS.

Test the failure casea wall that is not there

Hold the insert off the skin. The estimator should return a wall of near zero and the gate should refuse the take. If it returns a confident number instead, the method is wrong and you should say so.