Project DelphiTensors Workshop

Take-home E — low rank does not mean clean

Cut a real voice recording into short overlapping windows, take the STFT, truncate its SVD, invert, and measure SNR against the clean reference. 4.949 seconds at 48 kHz gives a (513, 465) STFT, so full rank is 465. The noisy input measures 5.00 dB, as constructed.

k SNR Retained energy
2 2.29 dB 33.2% worse than the noise we started from
5 4.76 dB 51.1% still worse
20 8.48 dB 70.1%
40 9.08 dB 78.3% best — +4.08 dB, at 8.6% of full rank
80 7.68 dB 87.1% past the peak
465 5.00 dB 100% exactly the noisy input again

Both ends fail, for opposite reasons. And note: going from k=40 to k=80 retains more energy and produces worse audio, because the energy being added back is noise.

Retained energy is not the criterion. Low rank does not mean clean; it means small.