← Results viewer

Sound Model Comparison

Spectrograms of one car and one truck pass-by: a real recording, the old simulation and the current simulation.

Real recordingOld simulationCurrent simulation

Car, 54 km/h, 3.1 m to the side

Spectrogram of a real car pass-by: broadband, no stripes — Real car pass-by, AcousticScenes, 1.1 s of joined 100 ms frames

Spectrogram of the old synthetic car: curved comb stripes above 2.5 kHz and level steps

Spectrogram of the new synthetic car: smooth, no stripes or steps

Truck, 50 km/h, 6.7 m to the side

Spectrogram of a real truck pass-by: broadband, no stripes — Real truck pass-by, AcousticScenes, 1.1 s of joined 100 ms frames

Spectrogram of the old synthetic truck: strong curved comb stripes and level steps

Spectrogram of the new synthetic truck: smooth, no stripes or steps

Microphone 0, 100 to 8000 Hz, same settings in every panel. Simulated clips are 6 s long; the vehicle passes at 3 s. The real recording is a different vehicle of the same type. It is only 1.1 s long (the dataset keeps 0.1 s of audio per frame), so it has no audio player.

Changes

Old simulation. The ground reflection caused curved stripes between 2.5 and 7 kHz, and the filter bank caused level steps in the spectrum.

Current simulation. No ground reflection, a smooth spectrum, and an independent signal for each sound source of a vehicle.

Real recordings. Neither the stripes nor the steps appear in the AcousticScenes recordings.

More examples of the old simulation

Six rows of spectrograms. Left: old synthetic pass-bys of cars and trucks, many with curved comb stripes. Right: real AcousticScenes pass-bys, without stripes. — Old simulation with ego-car noise: three cars and three trucks, next to real AcousticScenes pass-bys.