Room and reverberation: what works, and what does not

From a clap or a sweep it works. The result equals the app: the largest difference in reverberation time was 0 seconds. A simulated clap (a clap spectrum with room noise) against the bare impulse response, matched room for room, gave a median error of 3.9% over 155 rooms, with 86% within 10% and 98% within 20%. At high confidence (at least 35 dB of decay) 278 band measurements had a median error of 1.5%, with 91% within 10%. Against an independent numpy implementation 145 of 167 impulse responses could be read, with a median deviation of 0.5% (0.3% at a range of at least 35 dB). The reverberation time per octave band is Validated, with a reference basis.

From speech it did not work well enough. The tail measure followed the measured reverberation time only partly: a rank correlation of 0.56 (95% interval 0.12 to 0.81), against a bar of 0.70 with a lower bound of 0.50. The other measure had 0.303 (95% interval -0.20 to 0.71): a relationship was not shown. The slope is not calibrated to reverberation time: the measured slope is 20% of the theoretical one at 0.18 s and 87% at 1.71 s. Both measures stay indicative, and we struck the sentence that said one of them measures the room itself.

The reverb score failed. Its rank correlation with the measured reverberation was 0.06 and its AUC 0.52, and it rises with noise. It was withdrawn everywhere, and it stays on /lenses as withdrawn.

So: from a clap or a sweep, AudioLab measures reverberation time per octave band. From speech it cannot.

Clap or sweep recordings of real rooms and measured impulse responses. A clap gives the octave bands from 500 Hz up; the measurement refuses the lower ones. An impulse response or a sweep gives every band from 125 Hz.

• A real phone with a real clap.
• Speech or music instead of a clap, and 125 and 250 Hz from a clap.
• ISO 3382 certification.
• What reverberation costs a recogniser: not published yet. It will be added here when its numbers are in our register.